INTELLIGENT NETWORK ASSISTED BY

COGNITIVE COMPUTING AND MACHINE LEARNING

Machine-Learning-Based Positioning: A Survey and Future Directions

Ziwei Li, Ke Xu, Haiyang Wang, Yi Zhao, Xiaoliang Wang, and Meng Shen

Abstract munity, school, and factory. In different indoor

scenarios, the demands for different LBS-based
Widespread use of mobile intelligent termi- applications are also different. For example, hospi-
nals has greatly boosted the application of tal administrators need to track dangerous goods
location-based services over the past decade. and special patients in real time, and provide
However, it is known that traditional loca- navigation services for patients. In a mall, indi-
tion-based services have certain limitations such viduals need to find where their cars are parked
as high input of manpower/material resources, and the location of their intended stores. Further-
unsatisfactory positioning accuracy, and complex more, shopping center managers need to obtain
system usage. To mitigate these issues, machine- customer flow analysis data and direct them to
learning-based location services are currently precision advertising and marketing. In order
receiving a substantial amount of attention from to ensure safety, an LBS system in prisons and
both academia and industry. In this article, we schools can set up electronic fences to prevent
provide a retrospective view of the research special persons from walking out of designated
results, with a focus on machine-learning-based areas. Applications of LBS have been widely used
positioning. In particular, we describe the basic in transportation hubs. The system can automati-
taxonomy of location-based services and summa- cally identify arriving passengers and push traffic
rize the major issues associated with the design of information. Nowadays, the rapidly developing
the related systems. Moreover, we outline the key LBS and smart home technology provide peo-
challenges as well as the open issues in this field. ple with more comfortable and convenient living
These observations then shed light on the possi- environments. In a lighting system, a lamp can
ble avenues for future directions. turn on when a person enters but turn off when
he/she leaves. When people are on their way
Introduction home, the smart home service system can intel-
During the past decade, with the rapid develop- ligently adjust the temperature control system in
ment and spread of the Internet of Things (IoT), the home according to their location information
cloud computing, mobile computing, and intelli- and daily habits. In a factory, people can use the
gent terminals, the application of location-based LBS system to monitor and track people or assets,
services (LBS) has attracted wide attention from such as automatic check-in and patrol recorder.
both academia and industry. Based on users' loca- LBS have been applied to all aspects of human
tion information, LBS providers can offer richer, life, and they also have their own characteristics in
faster, and more accurate services. different application areas.
New technologies have spawned many new Although the LBS system brings great conve-
applications, as shown in Fig. 1. In the outdoor nience to people, there are still some disadvantag-
environment, satellite-based positioning technolo- es that restrict its promotion and popularization,
gies can provide convenient location services for including the large initial infrastructure investment
people, such as Global Positioning System (GPS)- of the LBS system, complex system usage, insuffi-
based vehicle navigation and cargo tracking. LBS cient positioning accuracy, and so on. Currently,
can also provide nearby living entertainment infor- the most urgent issue to be solved is how to tune
mation and intelligent path navigation services the problem of large data collection and comput-
for individual users. When encountering an emer- ing, and high manpower and material resources
gency, based on people location information, the cost. Artificial intelligence (AI) and machine learn-
public service department can provide emergen- ing (ML) have developed rapidly in recent years,
cy counseling and medical assistance services. and have been employed by many researchers
However, in the indoor environment, due to to improve LBS systems. We have noticed that
serious object occlusion and multipath effects some related research on deep learning and trans-
of signal propagation, satellite-based positioning fer learning concentrate on optimizing location
technologies face great challenges and cannot technology, demonstrating that the ML technol-
meet people’s daily demands. Since the 1990s, ogy provides a new opportunity to address these
indoor location technology has been the focus of challenges [1].
academic and industrial research, leading to a lot In this article, we present a survey and future
of achievements. Eight typical indoor scenarios directions of ML-based positioning. We present
are described in Fig. 1, including hospital, mall, the main localization technologies and outline
exhibition center, prison, transportation hub, com- the solutions currently available. Then, in a later

Digital Object Identifier: Ziwei Li, Ke Xu, Yi Zhao, and Xiaoliang Wang are with Tsinghua University; Haiyang Wang is with the Univeristy of Minnesota;
10.1109/MNET.2019.1800366 Meng Shen is with the Beijing Institute of Technology.

96 0890-8044/19/$25.00 © 2019 IEEE IEEE Network • May/June 2019

FIGURE 1. An overview of LBS, including both indoor and outdoor positioning scenarios.

section, we dwell on the newly developed deep ground control centers and 30 satellites. For these
learning and transfer learning research, with a spe- 30 satellites, the number of working satellites is
cial focus on the applications of positioning tech- 27, and the rest are spare satellites.
nology. Finally, we point out possible directions
and challenges of future studies. Indoor LocALIZAtIon
Outdoor positioning technology has brought
LocALIZAtIon great convenience to human life. However, the
In this part, we overview existing localization tech- accuracy of outdoor positioning still needs to be
nologies, which can be divided into two direc- improved. Especially in the indoor environment,
tions: outdoor localization and indoor localization. due to occlusion, uncertainty, and complicated
structure, existing outdoor positioning technolo-
outdoor LocALIZAtIon gies cannot meet the indoor positioning require-
With the rapid development of the satellite posi- ments. Therefore, a large number of researchers
tioning technology, outdoor LBS have gradually have begun to focus on indoor positioning tech-
penetrated into every aspect of people's lives, nologies.
bringing great convenience to people. The main- Existing indoor positioning technologies are
stream satellite positioning systems mainly include based on different signals, such as ultrasonic,
GPS, the Global Navigation Satellite System ultra-wideband (UWB), radio frequency identifica-
(GLONASS), BeiDou Navigation Satellite System tion (RFID), Bluetooth, and WiFi.
(BDS), and Galileo satellite navigation system Ultrasonic-based positioning technology main-
(GALILEO). First of all, GPS is a radio navigation ly uses reflective ranging, that is, transmitting ultra-
and positioning system developed on the Unit- sonic waves and receiving echoes generated by
ed States Navy navigation satellite system. It can the measured object, and then calculating the
provide users with geographical location on the distance between them according to the time dif-
Earth as well as time information. Another exam- ference, and finally determining the position of
ple is GLONASS, which is a second-generation the measured object. Ultrasonic-based position-
military satellite navigation system independently ing has high positioning accuracy, but it cannot
developed and controlled by the former Soviet penetrate walls and some obstacles. The ultrason-
Ministry of Defense. It is the second global sat- ic-based localization system faces the multipath
ellite navigation system after GPS, and consists effect, and it requires a large amount of invest-
of three parts: satellite, ground monitoring sta- ment in the underlying hardware facilities. Kim et
tion, and user equipment. In terms of technology, al. [3] proposed an ultrasonic reflections-based
GLONASS has better anti-interference ability than positioning system that exploits time difference of
GPS, but its single point positioning accuracy is arrival (TDoA) technology in location estimation.
not as good as GPS. BDS is a new global satellite Although it can achieve high positioning accu-
navigation system developed in China. It is com- racy with errors smaller than 35 cm, it deployed
posed of a space terminal, a ground terminal, and 20 beacons in the 620  980 c m2 room. UWB
a user terminal. BDS has been applied in many transmits data by sending and receiving nano-
fields, including surveying, telecommunications, second ultra-narrow pulses. It has the advantag-
emergency management, and so on. In Europe, es of strong penetrating power, high safety, and
GALILEO is more popular, which consists of two low system complexity. However, it is difficult to

IEEE Network • May/June 2019 97

 Deep learning Transfer learning stage, the researcher needs to obtain a certain
amount of WiFi signal strength information at the
Outdoor High scene stability High scene stability sampling points specified in the area to be locat-
localization Less labeled data Fewer similar scenes ed, forming the WiFi fingerprints in this area. Then,
in the second stage, matching the WiFi fingerprint
Indoor Large amount of labeled data Multiple similar scenes collected by the intelligent terminal in real time
localization Low scene stability Low scene stability with the offline fingerprint database takes place
to determine the user’s current actual location.
TABLE 1. Advantages and disadvantages of deep learning and transfer learning Current research in this stage mainly focuses on
for positioning. algorithm optimization, such as fingerprint match-
ing method, reduction sampling frequency, and
other auxiliary positioning technologies. At the
same time, in the offline stage, matters such as the
specific methods of selecting fingerprints, sam-
pling point selection, and prediction sample value
of wave function have also received extensive
attention.

mAchIne LeArnIng for Lbs

In recent years, with the rapid development of
ML, ML-based positioning technologies are also
very popular and have achieved good results. In
these ML-based positioning technologies, deep
learning and transfer learning are more practical
due to the complexity and diversity of the posi-
tioning environment. As shown in Table 1, we
summarize advantages and disadvantages of deep
learning and transfer learning for positioning. In
this part, we also summarize some typical work
related to deep learning and transfer learning.

deeP LeArnIng for PosItIonIng

Deep learning is a branch of ML. The motiva-
FIGURE 2. Deep learning for positioning. tion is to build and simulate a neural network
for human brain analysis and learning. It mimics
achieve large-scale indoor coverage with UWB, the mechanism of the human brain to interpret
and higher system construction costs limit the data. The concept of deep learning stems from
development of UWB-based positioning system. the research on artificial neural networks. The
RFID can automatically recognize a target multi-layer perceptron with multiple hidden layers
object through RF signal and acquires relevant is a deep learning structure. Deep learning com-
non-contact data. In general, RFID-based posi- bines low-level features to form more abstract
tioning technology has advantages of non-contact high-level representation attribute categories or
operation and non-line-of-sight transmission of sig- features to discover distributed feature represen-
nals, while it requires the deployment of separate tations of data. Deep learning has a wide range
devices, usually including reader, tag, host, and of applications in many areas, including computer
so on. Huang et al. [3] proposed a real-time RFID vision, speech recognition, natural language pro-
indoor positioning system based on Kalman filter cessing, search engines, finance, online advertis-
(KF) drift removal and Heron bilateration loca- ing, and more [4, 5].
tion estimation. It implemented an active RFID As shown in Fig. 2, in recent years, deep
tag, a portable RFID indoor positioning device, learning has also been attempted in wireless-sig-
and a terminal for localization and orientation nal-based localization algorithms to improve local-
indications. Bluetooth-based positioning was per- ization accuracy and reduce labor costs [6, 7].
formed by measuring Bluetooth signal strength or Wang et al. [8] proposed a deep-learning-based
fingerprinting methods. The Bluetooth chip has indoor fingerprinting system for indoor positioning
the advantages of small size, low power consump- named DeepFi, using channel state information
tion, and easy integration and deployment of the (CSI). This system is similar to the usual indoor
device. However, in a complex environment, the fingerprinting system, which includes an offline
Bluetooth signal is susceptible to interference training phase and an online localization phase.
from external noise signals. Deep learning is used to train all weights of a
WiFi, as the most widely deployed indoor deep network as fingerprints in the offline training
wireless network infrastructure, has covered phase. In order to reduce the system complexity,
most public places including shopping malls, it used a greedy learning algorithm. And then, in
museums, airports, railway stations, libraries, the online localization phase, to obtain accurate
and so on. Indoor positioning technologies position prediction, it employed a probabilistic
based on received signal strength of WiFi signal method based on the radial basis function.
have received the most extensive attention and Based on deep autoencoder, Khatab et al.
achieved successful application. This localization [9] proposed a new method to obtain high-level
system was divided into two stages from the algo- extracted features. It solved the shortcomings of
rithm aspect, that is, the offline data acquisition traditional methods that could not describe the
stage and online positioning stage. In the offline high-dimensional features of data. Furthermore,

98 IEEE Network • May/June 2019

taking into account the dynamic nature of the
environment, it can incrementally learn and con-
tinuously use new data to make the method more
stable and more accurate.

Transfer Learning for LBS

Transfer learning is another kind of ML method.
It is an optimization method for people to learn
things in new fields by comparing what they have
learned. In other words, transfer learning is more
suitable for new scenarios without much data. In
terms of data scarcity, there is deep reinforcement
learning [1]. The difference is that deep reinforce-
ment learning is suitable for gradually generating
data through interaction, and gradually finding the
optimal solution. For example, when we learn the
backstroke, it is easy to learn freestyle. Running
skills can be applied to race walking. As shown in
Fig. 3, transfer learning effectively reduces the ini-
tial training cost and reapplies the training model
from one task to another related task.
At present, supervised learning has been widely
used in business. However, supervised learning
needs to be based on massive training datasets. FIGURE 3. Transfer learning for LBS.
Through the correspondence between the existing
input data and the output data, a mapping function
model is generated by training, and the unknown reduce the need for fingerprint collection work,
data samples are predicted with the known model. significantly improving the positioning accuracy
This requires sufficient training samples, and the under the support of a small number of test sets.
training samples and test samples are distribut- It is conceivable that the main obstacle hindering
ed independently. In some scenarios, the cost of the large-scale promotion of LBS could be solved
obtaining data samples is very high, or it is impossi- with the support of transfer learning technology.
ble to obtain enough data samples at all. At present, many studies have achieved corre-
In recent years, transfer learning has been wide- sponding results in this respect.
ly studied in the field of ML. The advantage of Chang et al. [11] proposed FitLoc, a fine-
transfer learning is that we can apply the model to grained and low-cost device-free localization (DfL)
similar problems and get good results by making approach that can localize multiple targets in var-
minor adjustments to a trained model. For exam- ious areas. This research is based on compressive
ple, in the user evaluation of some new systems, sensing theory, using transfer learning to unify the
the data available for training in the target domain radio map over various areas. FitLoc clearly reduc-
is very little, which is usually unlabeled text. We es the deployment cost. Compared to other ways
need to analyze the emotional categories of user of using the Radio Topology Imagine model, Fit-
evaluation through natural language processing. It Loc uses an innovative transfer learning scheme
is difficult to deal with these texts using the tradi- during the localization phase. In order to reuse
tional ML method. By using transfer learning, it can the radio map of one area in various areas, FitLoc
analyze the user evaluation historical data of exist- projects the received signal strength (RSS) into a
ing similar systems, build the evaluation emotional subspace where the distribution distances over
analysis model, and then apply the model to the different areas are minimized.
user evaluation analysis of the new system. Trans- Chang's study is based on the isomorphic local-
fer learning can transfer models suitable for large ization system. Compared to the former, Zheng et
amounts of data to small amounts of data through al. [12] researched the cold-start heterogeneous
deep mining of common features of problems. It is device localization problem. in their paper, they
similar to label-less learning [10], which fundamen- trained the localization model from the data of
tally solves the problem of ML with a small amount the gauge device and tested it using the data from
of labeled sample data. the heterogeneous target device. In order to start
Typically, for an LBS system the amount of the positioning system in a cold-start environment,
data in the source domain is sufficient, whereas they wanted to find a robust representation of the
the amount of data in the target domain is small. feature. A new high-order pairwise (HOP) feature
This scenario is very suitable for transfer learning. representation and a novel constrained restricted
When we want to provide indoor LBS for one Boltzmann machine (RBM) model were proposed
new area, the fingerprint data is usually serious- in this paper. To gain more discriminative HOP
ly insufficient, while there is a large amount of characteristics and implement local deployment,
relevant fingerprint data available for training. this system combines the robust feature learning
There are some characteristic distribution differ- with localized model training.
ence between training data and implementing
data, including spatial distribution, wireless signal A Possible Vision and Challenges
hotspot layout, and so on. In this case, if a suitable The rapid development of LBS has facilitated peo-
transfer learning method is adopted, the wire- ple’s work and lives. Both public and business
less signal fingerprint accumulated in the source users have a wide range of demands for location
domain will transfer to a new space. It will greatly information and LBS. However, existing outdoor

IEEE Network • May/June 2019 99

 FIGURE 4. A possible vision for an LBS system.

and indoor LBS systems lack uniform standards. different scenarios, as well as the situational infor-
Different operators have different system architec- mation when people request the LBS.
tures and cannot achieve seamless connection.
Technology is always evolving rapidly with demand Challenges
in order to provide better LBS for people, which From a long-term perspective, we identify the fol-
requires not only in-depth academic research, but lowing challenges.
also the combined development of industry. How to Improve the Effect of Machine Learn-
ing Effectiveness under RSSI Changes: As we all
A Possible Vision know, received signal strength indication (RSSI)
Many researchers, engineers, and enterprises is greatly affected by changes in the surrounding
have been committed to the research and devel- environment, different population densities, room
opment of LBS, and people hold different opin- temperatures, humidity, and temporary large
ions on future visions. In this article, we describe a obstructions. How to represent the RSS fluctua-
possible vision of the future, integration of hetero- tion caused by environment change affects the
geneous LBS systems and seamless connection positioning accuracy and the algorithm validity.
between indoor and outdoor environments, as a Moreover, it is impossible to ensure that the RSSI
possible development direction. is collected in different physical spaces complet-
In Fig. 4, the right panel demonstrates a variety ed by the same equipment in practical applica-
of heterogeneous LBS systems that are suitable tion. This inevitably leads to the requirement of
for indoor and outdoor environments. In our daily normalization between existing fingerprint data
lives, we often need to navigate from a location in in different scenarios. Otherwise, the measure-
a large indoor environment to an outdoor destina- ment error between different devices may reduce
tion. For example, in an underground transporta- the positioning accuracy directly. The above two
tion hub, if we want to take a taxi to an exhibition problems are important to LBS system develop-
center, the current systems require us to use the ment, and should be solved effectively before
indoor LBS system of the transportation hub when large-scale commercial applications.
navigating to the taxi station, and then manually How to Avoid Negative Transfer: When RSS
switch to the outdoor positioning system when fingerprint data from a given region is transferred
navigating to the destination. to a new unmeasured region, it is important to
In the future development of LBS systems, we select the appropriate feature values and transfer
need to study not only the intercommunication objects. Due to differences in spatial layout and
between multiple scenes, but also the interconnec- wireless signal access point distribution among
tion between multiple heterogeneous LBS systems. regions, improper feature selection and transfer
There are many kinds of positioning technologies, mode may bring serious negative transfer effects,
which have different advantages and disadvan- resulting obviously decreasing positioning accu-
tages. Based on the ML method, optimizing and racy. How can we avoid negative transfer effec-
intergrating multiple heterogeneous LBS systems tively? There are two research directions. One
can effectively improve the positioning accuracy is to design an effective evaluation mechanism
and system robustness, and reduce the overall for transfer, judging the correlation between dif-
investment in LBS system solutions. Furthermore, ferent regions and providing appropriate feature
more in-depth and accurate LBS need to integrate selection that ensures the effectiveness of finger-
with a variety of networks, such as social networks, print feature transfer. Second, a negative trans-
industrial networks, and smart city networks. In fer discovery and correction algorithm should be
addition, edge cognitive computing can be used designed to evaluate the transfer effect in real
to save energy consumption[13]. People not only time, and improve by modifying or replacing the
desire location-based information, but also need feature mode to achieve satisfactory positioning
the semantic information of different locations in accuracy. It is believed that this problem will be

100 IEEE Network • May/June 2019

one important research direction of LBS technol- [3] C.-H. Huang et al., “Real- Time RFID Indoor Positioning Sys-
tem Based on Kalman-Filter Drift Removal and Heron-Bilat-
ogy in the future. eration Location Estimation,” IEEE Trans. Instrumentation and
How to Collect, Transfer, and Store a Massive Measurement, vol. 64, no. 3, 2015, pp. 728–39.
Amount of Location Data: ML-based positioning [4] A. M. Elkahky, Y. Song, and X. He, “A Multi-View Deep
needs large-scale high-accuracy location data as sup- Learning Approach for Cross Domain User Modeling in
Recommendation Systems,” Proc. Int'l. World Wide Web
port. In the LBS scenario, location data is generated Conf. Steering Committee, 2015, pp. 278–88.
with high frequency and large amounts. Therefore, [5] F. Monti et al., “Geometric Deep Learning on Graphs and
the collection, transmission, and storage of a mas- Manifolds Using Mixture Model CNNs,” Proc. CVPR, vol. 1,
sive amount of location data face great challenges. no. 2, 2017, p. 3.
[6] Y. Wang, K. Wu, and L. M. Ni, “Wifall: Device-Free Fall
In the research on location data collection, many Detection by Wireless Networks,” IEEE Trans. Mobile Com-
researchers obtain location data by crowd sourcing, puting, vol. 16, no. 2, 2017, pp. 581–94.
but there are certain errors between different sig- [7] L. Chen et al., “Human Behavior Recognition Using Wi-Fi
nal data acquisition devices. Moreover, much data CSI: Challenges and Opportunities,” IEEE Commun. Mag.,
vol. 55, no. 10, Oct. 2017, pp. 112–17.
transmission and storage face problems such as high [8] X. Wang et al., “CSI-Based Fingerprinting for Indoor Localiza-
energy consumption, high latency, and poor reliabil- tion: A Deep Learning Approach,” IEEE Trans. Vehic. Tech.,
ity. In order to promote the rapid development of vol. 66, no. 1, 2017, pp. 763–76.
positioning technology, it is necessary to solve the [9] Z. E. Khatab, A. Hajihoseini, and S. A. Ghorashi, “A Finger-
print Method for Indoor Localization Using Autoencoder
above problems in the future. Based Deep Extreme Learning Machine,” IEEE Sensors Let-
How to Ensure User Privacy and Facilitate Data ters, vol. 2, no. 1, 2018, pp. 1–4.
Sharing: Location information can reflect the user [10] M. Chen et al., “Label-Less Learning for Traffic Control in
behavior and thus be utilized in advertising and rec- an Edge Network,” IEEE Network, vol. 32, no. 6, Nov./Dec.
2018, pp. 8–14.
ommendation systems [14]. But the user behavior [11] M. Chen et al., “A Dynamic Service-Migration Mechanism
data contains personal sensitive information [15], in Edge Cognitive Computing,” ACM Trans. Internet Technol-
which may bring potential risks to the user if the ogy, vol. 19, no. 2, Apr. 2019, Article 30.
data is hacked by an attacker. Therefore, users have [12] L. Chang et al., “FitLoc: Fine-Grained and Low-Cost Device-Free
Localization for Multiple Targets Over Various Areas,” IEEE/ACM
a skeptical attitude toward existing LBS systems, Trans. Networking, vol. 25, no. 4, 2017, pp. 1994–2007.
which inhibits the promotion and development of [13] M. Chen et al., “A Dynamic Service-Migration Mechanism
LBS to some extent. In addition, LBS providers need in Edge Cognitive Computing,” ACM Trans. Internet Technol-
users’ accurate location data to provide richer, con- ogy, vol. 19, no. 2, Apr. 2019, article 30.
[14] Y. Zhao et al., “TDFI: Two-Stage Deep Learning Framework
venient, and accurate LBS support. As the existing for Friendship Inference via Multi-Source Information,” Proc.
user location privacy protection and data sharing IEEE INFOCOM, 2019.
model cannot solve the problem, a service model is [15] L. Hu et al., “Proactive Cache-Based Location Privacy Pre-
urgently needed, which can not only effectively pro- serving for Vehicle Networks,” IEEE Wireless Commun., vol.
25, no. 6, Dec. 2018, pp. 77–83.
tect user location privacy, but also encourage users
to share their non-sensitive location information. Biographies
Conclusion Ziwei Li received his B. Eng. degree and M. Eng. degree from
the Department of Information Science and Technology at Bei-
In this article, we present a survey and future direc- jing Forestry University, China, in 2011 and 2013, respectively.
tions of the LBS and positioning technologies, espe- Currently, he is pursuing a Ph.D. degree in the Department
of Computer Science and Technology at Tsinghua University,
cially from an ML perspective. Some positioning Beijing, China. His research interests include machine learning,
technologies based on different signals, such as sat- wireless networks, and mobile computing.
ellite, ultrasonic, UWB, RFID, Bluetooth, and WiFi,
are summarized. We then present the new research Ke Xu received his Ph.D. from the Department of Computer Sci-
ence and Technology of Tsinghua University, where he serves
achievements and applications of ML-based posi- as a full professor. He has published more than 100 technical
tioning. We believe that possible directions of posi- papers and holds 20 patents in the research areas of next gener-
tioning technology are integration of heterogeneous ation Internet, P2P systems, the Internet of Things, and network
LBS systems and seamless connection between virtualization and optimization. He is a member of ACM and has
guest edited several Special Issues of IEEE and Springer journals.
indoor and outdoor environments. There are still
many challenges that can be further explored in the Haiyang Wang received his Ph.D. degree from Simon Fraser
future. LBS applications have great potential and University, Burnaby, British Columbia, Canada, in 2013. He is
broad prospects, and ML will play an important role. currently an associate professor with the Department of Com-
puter Science, University of Minnesota at Duluth. His research
interests include cloud computing, peer-to-peer networking,
Acknowledgment social networking, big data, and multimedia communications.
The work in this article was in part supported by
the National Key R&D Program of China under Yi Zhao [S] received his B. Eng. degree from the School of Soft-
ware and Microelectronics, Northwestern Polytechnical Universi-
Grant No. 2018YFB0803405, China National ty, Xi'an, China, in 2016. Currently, he is pursuing a Ph.D. degree
Funds for Distinguished Young Scientists under in the Department of Computer Science and Technology at Tsin-
Grant No. 61825204, the Beijing Outstand- ghua University. His research interests include machine learning,
ing Young Scientist Project, the National Nat- social networks, and security. He is a student member of ACM.
ural Science Foundation of China under Grant Xiaoliang Wang received his Ph.D degree in computer science
61602039, and the Beijing Natural Science Foun- from Tsinghua University in 2017. Currently he is a postdoctoral
dation under Grant 4192050. Ke Xu is the corre- researcher in the Department of Computer Science and Tech-
sponding author of this article. nology at Tsinghua University. His research interests include
wireless networks and wireless sensor networks.

References Meng Shen [M] received his B.Eng degree from Shandong Uni-
[1] M. Mohammadi et al., “Semisupervised Deep Reinforcement versity, Jinan, China in 2009, and his Ph.D. degree from Tsinghua
Learning in Support of IoT and Smart City Services,” IEEE University in 2014, both in computer science. Currently, he serves
Internet of Things J., vol. 5, no. 2, 2018, pp. 624–35. at Beijing Institute of Technology as an associate professor. His
[2] K.-W. Kim et al., “Accurate Indoor Location Tracking Exploit- research interests include privacy protection of cloud-based ser-
ing Ultrasonic Reflections,” IEEE Sensors J., vol. 16, no. 24, vices, network virtualization, and traffic engineering. He received
2016, pp. 9075–88. the Best Paper Runner-Up Award at IEEE IPCCC 2014.

IEEE Network • May/June 2019 101