You are on page 1of 9


Intelligent 5G: When Cellular Networks

Meet Artificial Intelligence
Rongpeng Li, Zhifeng Zhao, Xuan Zhou, Guoru Ding, Yan Chen, Zhongyao Wang, and Honggang Zhang

Abstract times more aggregate throughput [2] and 10

times more at the link level [3] from 2010 to
5G cellular networks are assumed to be the 2020. Second, an increasing number of objects
key enabler and infrastructure provider in the are being digitalized to form the Internet of
ICT industry, by offering a variety of services Things, posing more stringent requirements on
with diverse requirements. The standardiza- latency, battery lifetime, etc. [4]. Therefore, to
tion of 5G cellular networks is being expedit- enhance service provisioning and satisfy the
ed, which also implies more of the candidate coming diversified requirements, it is necessary
technologies will be adopted. Therefore, it is to revolutionize the cellular networks with cut-
worthwhile to provide insight into the candi- ting-edge technologies. The standardization of
date techniques as a whole and examine the next-generation (5G) cellular networks is being
design philosophy behind them. In this article, expedited, which also implies more of the
we try to highlight one of the most fundamental candidate technologies will be adopted. This
features among the revolutionary techniques naturally raises questions such as which new
in the 5G era, i.e., there emerges initial intel- technologies might 5G cellular networks pos-
ligence in nearly every important aspect of sess, and which features will these technologies
cellular networks, including radio resource man- have in common?
agement, mobility management, service pro- From the very beginning, 5G cellular net-
visioning management, and so on. However, works were assumed to be the key enabler and
faced with ever-increasingly complicated con- infrastructure provider in the ICT industry, by
figuration issues and blossoming new service offering three types of services from enhanced
requirements, it is still insufficient for 5G cellular mobile broadband (eMBB) with bandwidth-con-
networks if it lacks complete AI functionalities. suming and throughput-driving requirements
Hence, we further introduce fundamental con- to new services such as ultra-reliable low laten-
cepts in AI and discuss the relationship between cy service (URLLC) and massive machine-type
AI and the candidate techniques in 5G cellular communications (mMTC). In that regard, though
networks. Specifically, we highlight the oppor- technologies such as densified cells and mas-
tunities and challenges to exploit AI to achieve sive multiple-input multiple-output (MIMO)
intelligent 5G networks, and demonstrate the are essential to boost capacity in the 5G era,
effectiveness of AI to manage and orchestrate it is cost-ineffective to deploy such techniques.
cellular network resources. We envision that Instead, 5G cellular networks mainly revolution-
AI-empowered 5G cellular networks will make ize themselves by initially embracing the intelli-
the acclaimed ICT enabler a reality. gence to agilely boost both spectrum efficiency
(SE) and energy efficiency (EE). Specifically, 5G
Introduction cellular networks provide alternative options for
Currently, fourth-generation (4G) cellular net- radio resource management (RRM), mobility
works are being globally deployed to provide management (MM), management and orchestra-
all-IP (Internet Protocol) broadband connectivi- tion (MANO), and service provisioning manage-
ty. Recalling that second-generation (2G) global ment (SPM) mechanisms. Hence, it is no longer
networks for mobile communications (GSM), necessary to build dedicated networks for indi-
debuted in 1991, just started to provide digital vidual services (e.g., the GSM-Railway commu-
voice telephony, and third-generation (3G) cel- nication networks). On the contrary, as depicted
lular networks, launched in 2001, initially pro- in Fig. 1, due to the development of smarter 5G
vided mobile Internet solutions. It took less than networks, it will be feasible to provide custom-
30 years to successfully transform cellular net- ized end-to-end network slices (NS) [5] to simul-
works from pure telephony systems to networks taneously satisfy distinct service requirements,
that can transport rich multimedia content [1] such as ultra-low latency in URLLC and ultra-high
and have a profound impact on our daily life. throughput in eMBB.
Nowadays, the landscape of the information There is no doubt that 5G cellular networks
communication technology (ICT) industry is rap- will tailor the provisioning mechanisms for differ-
idly changing. First, mobile broadband access is ent predefined services and pave the way for the
expected to have a drastic increase with 1000 application of complete intelligence. However,

Rongpeng Li, Zhifeng Zhao (corresponding author), and Honggang Zhang are with Zhejiang University; Xuan Zhou and Yan Chen are with Digital Object Identifier:
Huawei Technologies Co., Ltd.; Guoru Ding is with PLA University of Science and Technology; Zhongyao Wang is with the Alibaba Group. 10.1109/MWC.2017.1600304WC

IEEE Wireless Communications • October 2017 1536-1284/17/$25.00 © 2017 IEEE 175

networks, and transform cellular networks from
4G: all-IP Voice, being network-centric to being user-centric and
broadband 4G cellular networks
text, information-centric with significant SE and EE
connectivity mobile Internet improvement.
via phone
Radio Resource Management
Current 4G cellular networks heavily rely on
orthogonal frequency-division multiplexing
mobile Internet,
~20 Gb/s ... (OFDM) as the signal bearer and the base of
associated access schemes. Since OFDM can be
used in both frequency-division duplex (FDD)
URLLC, Automobile, and time-division duplex (TDD) formats, FDD
1 ms 5G cellular networks factory,
and TDD 4G cellular networks share a similar
frame structure by grouping a static number of
subcarriers and symbols into one resource block
mMTC, (RB). Benefiting from the satisfactory subcarrier
200,000/km 2 orthogonality in OFDM, information transmitted
climate, in different RBs can be separately decoded at
... the receivers with limited computational cost.
However, it is stubborn to use OFDM to simulta-
FIGURE 1. 5G cellular networks: a key enabler to all mobile devices across all neously satisfy service requirements from differ-
industries. ent users with various channel conditions, user
terminal (UE) capabilities (multiple access sup-
port, full duplex mode, feature or smart phones),
it is still challenging and time-consuming for 5G mobility, frequency bands, and so on. Given
cellular network operators to solve ever-increas- that, 5G cellular networks aim to introduce new
ingly complicated configuration issues and satisfy waveforms and provide softer air interfaces.
evolving service requirements, since 5G cellular Specifically, filter-bank multi-carrier (FBMC) and
networks merely possess more technical options unified-filter multi-carrier (UFMC) are famous
to deal with predefined intelligent problems, rath- candidates for more flexible frame structures and
er than gain the ability to interact with the envi- waveforms in the 5G era. As their names imply,
ronment (e.g., traffic load, service characteristics). FBMC and UFMC both add filters to combat out-
Fortunately, such an interaction falls into the field of-band leakage across subcarriers and make it
of artificial intelligence (AI), which is dedicated unnecessary to strictly synchronize across RBs.
to empowering machines and systems with intel- Therefore, 5G cellular networks can provide
ligence similar to that of humans. Hence, it is different air interface solutions in different RBs,
promising to apply AI to 5G cellular networks to in which different multiple access schemes, TTI
deal with newly emerging issues. (transmission time interval) parameters, wave-
In this article, we will try to answer what key forms, and duplex mode, pilot signals, etc., can
technical progress is in 5G cellular networks, why be well defined [7]. For example, as seen in Fig.
it is crucial to embrace AI in the 5G era, and how 2a, larger bandwidth and symbol length can be
AI can contribute to management and orchestra- applied to eMBB to yield a higher rate, while
tion in the 5G era. smaller TTI can be configured for URLLC to
shorten response latency.
The Intelligence in Similar to the evolution from OFDM to FBMC/
UFMC, 5G cellular networks potentially adopt
5G Cellular Networks non-orthogonal multiple access (NoMA) schemes
AI is the science and engineering of making such as sparse coding multiple access (SCMA).
machines as intelligent as humans, and has long Such NoMA schemes overlap information from
been applied to optimize communication net- two transmitters in the same radio resource and
works in diverse configurations [6]. According apply successive interference cancellation (SIC)
to the extent of intelligence, AI could be divided receivers (or even more computationally-exhaus-
into two levels. The first and basic level of AI is tive maximum-likelihood receivers) to decode the
that one machine or entity can provide multiple received information. Apparently, NoMA could
pre-defined options and respond to the environ- potentially lead to higher throughput. Moreover,
ment in a different yet deterministic manner. For another advantage of NoMA is that it makes pos-
example, as discussed later, 5G will allow granted sible grant-free transmission in the uplink (UL), if
and grant-free transmission for eMBB and mMTC the UE identity and the preamble for grant-free
services, respectively. In other words, the network UL transmission are mapped together. Instead of
will intelligently adjust the configuration after waiting for resource allocation commands as in
detecting different pre-defined service indicators. 4G cellular networks, it is feasible to decode the
The second and complete level of AI is that one overlapped information from two UEs at the same
machine or entity possesses full capability to inter- resources by using SIC receivers. From Fig. 2b, in
act (e.g., sense, mine, predict, and reason) with spite of the reliability advantage for granted trans-
the environment. More importantly, the machine mission, grant-free transmission in UL could avoid
or entity is able to learn how to make appropri- the cumbersome signaling procedures and save
ate responses, even when it faces strange sce- latency for small packets at a trivial performance
narios or tasks. In this section, we will highlight loss. Comparatively, 5G cellular networks even-
how the candidate technologies grant preliminary tually have one alternative option, which is quite
intelligence (i.e., the basic level of AI) to cellular suitable for mMTC service.

176 IEEE Wireless Communications • October 2017

Flexible eMBB BS mMTC
Random access (RA) preamble

Flexible RA response
symbol Message 3
Packets with preambles
length Contention resolution
resource allocation
Time Packets


Granted Grant-free
(a) (b)
MME Mobile edge
update Data networks
message packet

update Data Core
message packet networks


(c) (d)

FIGURE 2. Candidate technologies for intelligent cellular networks: a) flexible bandwidth and symbol length enabled by FBMC and
UFMC; b) granted and grant-free transmission enabled by NoMA; c) flexible mobility management schemes; d) dynamic service
provisioning stack.

Mobility Management Internet, and initiated the project called Central

In 4G cellular networks, there exist two states Office Re-architected as a Datacenter (CORD).
to manage UEs’ location awareness to the core CORD has successfully completed the virtualiza-
network (e.g., evolved packet system (EPS)). All tion of existing hardware devices such as CPE
EPS mobility management (EMM)-connected UEs (customer premises equipment), OLT (optical line
should periodically report their locations, so as to transmission) and BNG (broadband network gate-
guarantee the session continuum and informa- way), and produced software counterparts (e.g.,
tion reachability. Definitively, it is resource-con- applications running on open network operating
suming to treat all UEs the same. Instead, some system (ONOS)) on top of commodity hardware.
UEs at static positions (e.g., UEs for mMTC meter- Moreover, CORD has provided a framework on
ing services) should only need to report at the which these software elements (plus any other
very beginning of network attachment. Taking cloud services an operator may want to run) can
account of practical considerations, 5G cellular be plugged into, leading to a coherent end-to-end
networks introduce multiple-tier mobility manage- system. Therefore, operators of 5G cellular net-
ment states to make the mobility management works might borrow the concept of CORD and
mechanism more flexible. For example, for mMTC deploy selected functionalities according to their
UEs possessing characteristics such as immobil- own demands. Meanwhile, in order to orchestrate
ity, cost-sensitivity, and stringent requirements services from different vendors, 5G network oper-
on energy-efficiency, 5G cellular networks will ators can leverage a more centralized SDN con-
wait for the communication request from UEs troller and adopt various means such as exposing
and reactively start the data transmission (Fig. the same infrastructure-level interfaces or using
2c). Meanwhile, 5G cellular networks also tailor common cloud operating systems (e.g., ONOS) to
mobility management for some vertical industries, shield differences between multi-vendor hardware
based on regional characteristics. In other words, servers in a distributed deployment manner [5].
once UEs enter a specific region, they could be
granted higher-level support (e.g., dual connec- Service Provisioning Management
tivity) for mobility management and thus update In addition to softer air interface enabled by
their locations in a more proactive manner. FBMC or UFMC, it is also expected in 5G to
intelligently program the forwarding route of one
Management and Orchestration service by leveraging the application interfaces
Recently, the industry has witnessed the increas- (APIs) in SDN and have a more flexible service
ing maturity of software-defined networks (SDN) provisioning stack. With the evolution of SDN
[8]. In particular, some well known operators such and network function virtualization (NFV), 5G cel-
as AT&T, China Mobile, Telefonica, and vendors lular networks have advocated a revolutionary
such as Cisco and Huawei, have co-established concept called network slicing (NS) [9]. Instead
the Open Networking Lab ( to of building dedicated networks for different ser-
bring openness and innovation in SDN to the vices, NS allows operators to intelligently create

IEEE Wireless Communications • October 2017 177

Cellular networks have customized network pipes to provide optimized for network maintenance, it is essential for cellular
alternative options in solutions for different services that require diverse networks to observe environment variations, learn
the 5G era for access functionalities, performance metrics, and isolation uncertainties, plan response actions, and configure
and service provi- criteria. Moreover, session management in 5G the networks properly. Coincidentally, AI main-
will be able to adapt to UE attributes and service ly solves how to learn the variations, classify the
sioning mechanisms requirements, by adjusting configurations such as issues, forecast future challenges, and find poten-
and thus gain the session categories, anchoring points, and service tial solutions, by interacting with the environment.
foundation to apply continuum capabilities. Specifically, mobile edge Therefore, cellular networks could leverage the
preliminary intelligence. has potentially evolved to replace its forward- concept of cognitive radio [10] and interact with
However, 5G cellular ing-only functionality to an area equipped with the environment using AI, so as to fully accelerate
storage, memories, and computational power the evolution and enter into a brand-new intelli-
networks are still lag- capabilities [3]. Therefore, according to practical gent 5G era.
ging behind what is requirements, UEs could select anchoring and for- AI has evolved to multi-disciplinary techniques
actually required warding points between the anchoring point (e.g., such as machine learning, optimization theory,
in practice. serving gateway) in core networks (CNs) and the game theory, control theory, and meta-heuristics
mobile edge networks. For example, in Fig. 2(d), [11]. Among them, machine learning belongs to
services with stringent requirements on mobility one of the most important subfields in AI. Usually,
and service continuum could shift their anchoring depending on the nature of the learning objects
points to the edge networks with closer proximity. and signals to a learning system, machine learning
Moreover, in 4G, device-to-device (D2D) commu- is typically classified into three broad categories:
nication merely supports proximity services and Supervised Learning: A supervised learning
public safety communications. But network-assist- agent will be fed with example inputs and their
ed direct communication between vehicles and desired outputs, and aims to determine a general
UEs comes to a reality, and the vehicle-to-vehicle rule that nicely maps inputs to outputs. Super-
infrastructure (V2X) services are becoming a hot vised learning has been widely applied to solve
topic to better accommodate the URLLC services channel estimation issues in cellular networks. For
of vertical industries (e.g., automobiles). example, assume that there exists a wireless chan-
Thanks to the huge advance in signal process- nel h, the receiver tries to exploit the transmit pre-
ing capabilities evolved as Moore’s Law, 5G cellu- amble s and the received signal y = hs + n0 (with
lar networks can take advantage of advanced yet n0 denoting the noise) to estimate h. For such a
computation-consuming technologies in almost supervised learning problem, it is common to use
every aspect spanning from the physical layer to probabilistic models to characterize the transition
the network architecture. Therefore, 5G cellular probability P(y|s) from s to y and take advan-
networks are able to provide alternative options tage of the well known Bayes learning methods
for different scenarios, exhibit some preliminary to obtain the results. The well known Kalman fil-
intelligence, and satisfy the minimal requirements tering and particle filter methods also play a very
to adopt complete AI. important role in optimizing cellular networks.
Unsupervised Learning: Compared to the
Artificial Intelligence for aforementioned supervised learning, the input
information for unsupervised learning does not
Cellular Networks possess priori labels. Therefore, the unsupervised
Cellular networks have alternative options in the learning agent has to depend on its own capa-
5G era for access and service provisioning mech- bility to find the embedded structure or pattern
anisms and thus gain the foundation to apply in its input. Usually, unsupervised learning aims
preliminary intelligence. However, 5G cellular to discover hidden patterns and find the suitable
networks are still lagging behind what is actual- representation in the input data. In the field of AI,
ly required in practice. First, the number of con- unsupervised learning is applied to estimate the
figurable parameters in a typical 4G node has hidden layer parameters in neural networks and
increased to 1500 from 500 in a 2G node and plays an important role in deep learning meth-
1000 in a 3G node [4]. If this trend continues, a ods. Meanwhile, unsupervised learning may be
typical 5G node is expected to have 2000 or more the most widely applied AI category in cellular
parameters. Therefore, it is critical to enhance networks. For example, principal component
intelligence in the 5G era to realize the self-orga- analysis (PCA) and singular value decomposition
nizing features (e.g., self-configuration, self-opti- (SVD) methods have been used to manipulate the
mization, and self-healing). Second, the service receiving matrix of massive MIMO to reduce the
types (e.g., eMBB, URLLC, mMTC) defined in the computational complexity. Moreover, 5G NoMA
5G era are static. However, new types of services receivers also adopt some factor graph-based
continually evolve, and the pattern in existing ser- methods such as expectation-maximization and
vices frequently changes as well. In this case, 5G message-passing algorithms to achieve lower bit
cellular networks still lack functionalities to auto- error rate. On the other hand, some classifiers
matically recognize a new type of service, infer the such as the K-means Algorithm are also useful to
appropriate provisioning mechanism, and establish detect network anomalies.
the required network slice. Third, 5G cellular net- Reinforcement Learning: Inspired by both con-
works heavily depend on a centralized network trol theory and behaviorist psychology, the rein-
architecture in SDN, and still lack the agility and forcement learning agent could obtain its goal by
robustness under the scenario of ever-increasing interacting with a dynamic environment. Howev-
heterogeneous and complicated cellular networks. er, the agent does not have explicit knowledge
To self-organize parameters that become signifi- of whether it has come close to its goal. Instead,
cantly larger, auto-build the network slices for the agent should take actions in an environment
emerging services, and gain sufficient flexibility so as to maximize the cumulative reward in a

178 IEEE Wireless Communications • October 2017

Modules Examples Algorithms Comments

Sensing Detection of network Logistic Regression (LR) Hypothesis test plays an important role in this aspect. But different
anomalies or events by Support Vector Machine (SVM) algorithms have specific scenarios. Compared to SVM, LR is more suit-
multiple-entry data from Hidden Markov Model (HMM) able for sensing scenarios with a heavy number of property combinat-
hybrid sources ions and stringent accuracy requirements. On the other hand, HMM is
also applicable for sensing if we try to compute the state’s probability
and regard a comparably larger probability as the occurrence of
anomalies or events.

Mining Classifying services according Supervised learning: Supervised learning heavily relies on the labeling quality of data
to the required provisioning • Gradient Boosting Decision Tree (GBDT) samples, while unsupervised learning depends on the accuracy or
mechanisms (e.g., band- Unsupervised learning: suitability of parameter (e.g., threshold) settings.
width, error rate, latency) • Spectral Clustering
• One-class SVM
• Replicator Neural Networks (RNN)

Prediction Forecasting the trend of UE Kalman Filtering (KL) KL/ARMA/ARIMA could well follow the variations of a one-time se-
mobility or the traffic volume Auto-Regressive Moving Average (ARMA) quence, but fail to capture the characteristics behind this sequence. On
of different services Auto-Regressive Integrated Moving Average (ARIMA) the other hand, DL algorithms like RNN and LSTM have the capability
Deep Learning (DL): to find the embedded characteristics and leverage the long-time
• Recurrent Neural Networks (RNN) dependency in the sequence. Meanwhile, CS is a dedicated tool to
• Long-Short Term Memory (LSTM) investigate the universal sparsity in mobile traffic series and resources
Compress Sensing (CS) (e.g., BSs).

Reasoning Configuration of a series of Dynamic Programming (DP) DP, which might belong to a generalized sense of AI, is generally
parameters to better adapt • Branch-and-Bound Method exploited to solve the Bellman equation, based on complete knowl-
services. • Primal-and-Dual Method edge of the considered environment. In contrast, RL approximates the
Reinforcement Learning (RL) optimal solution of the Bellman equation without knowing the envi-
• Actor-critic Method ronment a priori, by iteratively updating its policy or value function.
• Q-Learning Method Besides, a combination of RL and TL could yield superior results.
Transfer Learning (TL)
TABLE 1. Typical AI algorithms to enhance cellular networks.

AI modules
4G 5G Intelligent 5G
Sensing Mining Prediction Reasoning

Services MBB eMBB/mMTC/URLLC Service-aware

Granted or grant-free
RRM Granted Flexible bandwidth ü ü û ü UE-specific on-demand
Flexible symbol length

MM Unified On-demand ü û ü ü Location tracking/awareness

Enhanced self-organizing and

MANO Simple Operator-tailored ü ü ü ü
trouble-shooting capability

Network slice auto-

SPM Unified End-to-end NS û ü ü ü

TABLE 2. The evolution toward intelligent 5G.

Markov decision process (MDP). Therefore, rein- UEs, base stations (BSs), and network entities.
forcement learning demonstrates strong pattern Table 2 illustrates functionalities upon which intel-
recognition ability. Researchers in the field of cog- ligent cellular networks may be built.
nitive radio usually model the dynamic transition Figure 3 illustrates a possible AI-empowered
of spectrum availability as a Markov chain, and 5G cellular network architecture, in which an
extensively apply reinforcement learning meth- AI controller will act as an application on top of
ods (e.g., Q-learning and the actor-critic method ONOS or an independent network entity, and
[12, 13]) to make the decision whether or not it is communicate with RAN, CN, or global SDN con-
suitable for secondary transmission in one primary trollers using open interfaces. Specifically, the AI
licensed spectrum, in terms of least interference center will read service-level agreements (e.g.,
to the primary spectrum. requirements on rate, coverage, failure duration,
Table 1 summarizes what typical AI algo- redundancy, etc.), UE-level information (e.g.,
rithms could solve. Apparently, AI can be used receiver category, battery limitation), network-lev-
to enhance the response of cellular networks to el information (e.g., spectrum, number of serving
stimuli by learning key network parameters. For subscribers, QoS (quality of service), key perfor-
example, AI makes it possible to sense in a timely mance indicators of network functions, scheduled
manner the variations in network traffic, resource maintenance period, etc.), and infrastructure-level
utilization, user demand, and possible threats, and information (e.g., server type, CPU, memory, stor-
further makes it possible to smartly coordinate age, network standard) from the SDN control-

IEEE Wireless Communications • October 2017 179


Global controller

AI center

... Service Traffic Requirement...

awareness prediction cognition
User Power
Scheduling Routing Caching Safety
association control

RAN controller CN controller

Data interface
... Traffic UE Resource ...
Pico and macro Shared signal Gateways Switches MMEs MiddleBoxes
BSs processing pool

FIGURE 3. 5G cellular networks enabled by AI.

lers, so as to get in touch with cellular network by provisioning different services and types of
data such as traffic information, UEs, and network UEs under various channels, network entity
resources. Afterward, the AI center will utilize its configurations, and energy consumption con-
embedded modules (e.g., sensing, mining, predic- ditions. In particular, AI could exploit cellular
tion, and reasoning) to process the obtained infor- network data to forecast potential events and
mation, and feedback learning results, which may predict traffic volume and help to pre-allocate
include traffic characteristic analysis reports (e.g., network resources. Meanwhile, AI provides a
service provisioning suggestion), UE-specific con- unified means to mine the relevancy in such
trolling information (e.g., serving priority, band- abundant data and helps build a more concrete
width allocation, mobility tracking command), and mapping from service requirement to network
network configuration notification (e.g., param- configuration. Furthermore, AI could generate
eter adjustment, access method, network error some operating reports to describe and sum-
alert), to the SDN controllers. For example, AI marize the subscriber and network experience
leverages the sensing module to track the loca- statistics, which is relevant when setting billing
tion of UEs and uses the predicting module to and market policies.
forecast the mobility trend based on the historical Inter-Networking of Heterogeneous Cellular
moving pattern. Afterward, it takes advantage of Networks: Currently, operators have deployed
the reasoning module and proactively notifies the heterogeneous BSs in the 4G era, including pico-
UEs to update the location record, so as to pre- cells (providing high capacity), micro-cells (pro-
pare handover resources and save signaling cost viding wide coverage for eMBB) and macro-cells
of mobility management. (supplying even wider coverage for signaling and
On the other hand, 5G cellular networks can mMTC services). AI could analyze the require-
maintain the normal working status under the ments of one emerging service and contribute
condition of potential damages (e.g., hacking) to to the selection of the most appropriate access
the AI center. Meanwhile, the AI center could point to accommodate such a service, in terms
(semi-)periodically exchange information with the of SE, EE, or other more complicated criteria. For
SDN controllers in normal states, while it starts example, AI could generate UE-specific policies
emergent responses to schedule the minimum to make some UEs attach to pico BSs for larger
required resources, once the conventional SDN throughput while letting some UEs connect to
controllers encounter malfunctions. Therefore, macro BSs to maintain fundamental information
compared to the complete centralized architec- exchange.
ture in conventional networks, the AI center and Difficulties in an Operator Supporting Sub-
the SDN controllers virtually constitute a multi-ti- system: Usually, cellular networks merely rely
er decision-making system, thus being able to on thresholds to monitor network anomalies.
improve the network robustness. Therefore, operator engineers have to be vigilant
enough to systematic alerts and read user guides
Opportunities to cope with unexpected network conditions. But
In addition to the benefits to the RRM, MM, AI could use cellular network data to derive com-
MANO, and SPM, AI could further contribute to mon network traffic patterns. Therefore, when net-
solving the following issues. works experience traffic with unfamiliar patterns,
Overloading of Cellular Network Data: Cel- AI can start troubleshooting at the very beginning.
lular networks generate vast volumes of records Similarly, networks could take advantage of AI to

180 IEEE Wireless Communications • October 2017

The BS transfer The BS
management management
in Period 1 in Period 2

Actor Exotic policy

Policy Actor
Overall policy
Native policy
Value function

Value function
TD error Action
BS 1: active
.. Action
State Cost .
BS i: sleeping BS 1: active
.. State Cost ..
. .
Environment: BS N: active BS i: sleeping
the base stations Environment: ..
the base stations BS N: active
Classical RL Algorithm
RL Algorithm with data

FIGURE 4. The AI application framework for reinforcement learning-based greener cellular networks.

shield against potential safety threats, once AI per- tive of the huge variations in traffic load. There-
ceives activity and address anomalies. fore, benefiting from cloud pooling of baseband
Challenges in Integrating RANs and CNs: Usu- resources, an SDN controller [8] can be leveraged
ally, the management of RANs and CNs is iso- to sense traffic variations and adjust the working
lated, thus being not scale enough for network status of under-utilized BSs, thus improving ener-
evolution. As mobile edge computing becomes gy efficiency. Meanwhile, an AI center plays a cru-
more common, AI grants the controller more cial role in learning traffic variations and adjusting
power and capability to jointly schedule wireless BS switching policy. Here, we briefly talk about
and wired resources, choose the appropriate con- two AI schemes to design traffic-aware greener
tent distribution and caching server (e.g., edge cellular networks, and show how AI could effec-
and core server), and provide more unified pro- tively solve this problem.
tection against possible network threats. The most intuitive approach is to first fore-
cast traffic loads in the near future and then
Challenges adjust the status of BSs, so as to satisfy the
In spite of the apparent opportunities, there are predicted traffic loads but incur minimal ener-
also challenges to apply AI to cellular networks. gy consumption. For traffic prediction, we
First, in the 5G era, network data is a double-edged can resort to the prediction module of AI. For
sword. It definitively provides precious oppor- example, our previous work [14] modifies the
tunities for AI to analyze trends and recognize popular ARMA algorithm by incorporating traf-
patterns. However, it is also difficult to derive a fic sparsity in both the temporal and spatial
simple model or pattern that perfectly matches domains and demonstrates the prediction error
the data. Therefore, the derived results, which for aggregate traffic records collected from
possibly consist of lots of parameters, are very dif- China Mobile could be as low as 15 percent,
ficult to read and lose value for practical applica- in terms of normalized root mean square error.
tion. Second, in order to save and process cellular Furthermore, we also predict the service-level
network data in a timely manner, a significant traffic with more bursty property in [15] and
amount of storage and computational resources demonstrate appealing accuracy as well, by
are needed, and there might be threats to infor- deriving the traffic model (e.g., a-stable mod-
mation security. Also, it usually is necessary to col- els) from realistic traffic records and utilizing a
lect data in a centralized manner before applying stable model-based compressive sensing algo-
most AI algorithms. These factors inevitably add rithm. 1 On the other hand, in order to deter-
to the computational capability of network enti- mine the appropriate BS switching policy, we
ties and BSs and put a huge burden on the practi- take advantage of the branch-and-bound algo-
cal cost of products. rithm, one kind of dynamic programming meth-
od, to solve the formulated optimization theory
Use Case: Traffic-Aware [14] and show that along with practical traffic
variations, significant energy savings could still
Greener Cellular Networks be expected. In particular, when traffic hits to
In this section, we demonstrate how to take the lowest point in the early morning (from 6
advantage of AI to enhance the MANO, so as AM to 8 AM), the AI-induced BS switching pol-
to build greener cellular networks [12]. It is well icy merely costs 55 percent of the energy that
known that over 80 percent of power consump- would be consumed if we emply no energy sav-
tion takes place in RANs, especially the BS, since ing scheme. 1 It can be regarded as a suc-
the present BS deployment is on the basis of peak In order to attain the BS switching solution, it is cess to combine the AI sens-
traffic loads and generally stays active irrespec- also viable to merge the prediction and reasoning ing and prediction modules.

IEEE Wireless Communications • October 2017 181

Based on the sensing, modules together, by applying a Markov chain to [4] A. Imran, A. Zoha, and A. Abu-Dayya, “Challenges in 5G:
How to Empower SON with Big Data for Enabling 5G,” IEEE
prediction, and reason- model possible traffic load variations and making Network, vol. 28, no. 6, Nov. 2014, pp. 27–33.
ing modules, AI could use of the actor-critic algorithm in Fig. 4 [13], a [5] X. Zhou et al., “Network Slicing as a Service: Enable Indus-
contribute to designing reinforcement learning (RL) approach [12]. Spe- tries Own Software-Defined Cellular Networks,” IEEE Com-
cifically, the AI center would first estimate traffic mun. Mag., vol. 54, no. 7, Jul. 2016, pp. 146–53.
traffic-aware greener [6] X. Wang, X. Li, and V. C. M. Leung, “Artificial Intelli-
load variations based on the on-line experience. gence-Based Techniques for Emerging Heterogeneous Net-
cellular networks. Afterward, it can select one of the possible BS work: State of the Arts, Opportunities, and Challenges,” IEEE
Meanwhile, such a switching operations under the estimated circum- Access, vol. 3, 2015, pp. 1379–91.
stance and then decrease or increase the prob- [7] C. L. I et al., “New Paradigm of 5G Wireless Internet,” IEEE
design represents the JSAC, vol. 34, no. 3, Mar. 2016, pp. 474–82.
typical methodology to ability of the same action to be later selected [8] R. Li et al., “The Prediction Analysis of Cellular Radio Access
according to the feedback cost information from Network Traffic: From Entropy Theory to Networking Prac-
exploit AI and reflects SDN controllers, by updating the policy and value tice,” IEEE Commun. Mag., vol. 52, no. 6, June 2014, pp.
the effectiveness of AI function. After repeating the actions and gradual- 238–44.
[9] 3GPP, “Study on architecture for next generation system (TR
on future cellular net- ly knowing the corresponding costs, the AI cen- 23.799),” 2016; available: http: //
works. ter would know how to switch the BSs for one port/23799.htm.
specific traffic load profile. Meanwhile, as cellular [10] J. Mitola and G. Q. Maguire, “Cognitive Radio: Making
network traffic exhibits strong self-similarity, the Software Radios More Personal,” IEEE Personal Commun.,
vol. 6, no. 4, Aug. 1999, pp. 13–18.
AI center could exchange the learned results in [11] J. Qadir et al., “Artificial Intelligence Enabled Networking,”
different periods and optimize the strategy in a IEEE Access, vol. 3, 2015, pp. 3079–82.
faster manner. Our numerical results in a practi- [12] R. Li et al., “TACT: A Transfer Actor-Critic Learning Frame-
cal BS deployment scenario with simulated traffic work for Energy Saving in Cellular Radio Access Networks,”
IEEE Trans. Wireless Commun., vol. 13, no. 4, Apr. 2014, pp.
traces also show that without knowing traffic vari- 2000–11.
ations beforehand, this RL method [12] could still [13] V. Konda and J. Tsitsiklis, “Actor-Critic Algorithms,” SIAM J.
converge quickly and approach the energy saving Contr. Optim., vol. 42, no. 4, 2000, pp. 1143–66.
performance of a solution with perfectly predict- [14] R. Li et al., “Energy Savings Scheme in Radio Access Net-
works via Compressive Sensingbased Traffic Load Predic-
ed traffic knowledge. tion,” Trans. Emerg. Telecommun. Tech. (ETT), vol. 25, no. 4,
In a word, based on the sensing, prediction, Apr. 2014, pp. 468–78.
and reasoning modules, AI could contribute to [15] R. Li et al., “The Learning and Prediction of Application-Lev-
designing traffic-aware greener cellular networks. el Traffic Data in Cellular Networks,” arXiv:1606.04778 [cs],
June 2016; available:
Meanwhile, such a design represents the typical
methodology to exploit AI and reflects the effec- Biographies
tiveness of AI on future cellular networks. R ongpeng L i received the Ph.D and B.E. from Zhejiang
University, Hangzhou, China and Xidian University, Xi‘an,
Conclusion China, in June 2015 and June 2010, respectively, both as
“excellent graduates.” He is now a postdoctoral researcher
In this article, we have suggested applying AI at Zhejiang University, Hangzhou, China. From August 2015
to cellular networks. We first discussed the ini- to September 2016, he was a researcher at the Wireless
tial intelligence emerging in nearly all aspects of Communication Laboratory, Huawei Technologies Co. Ltd.,
5G cellular networks, including radio resource Shanghai, China. His research interests focus on applica-
tions of artificial intelligence, data-driven network design,
management, mobility management, gener- and resource allocation of cellular networks (especially
al management and orchestration, and service full-duplex networks). He was granted by the National Post-
provisioning management. Following such intel- doctoral Program for Innovative Talents, which had a grant
ligence, we argued it is still essential to bring ratio of 13 percent in 2016.
more AI functionalities to 5G cellular networks
Zhifeng Zhao (corresponding author) is an associate professor
by envisioning several prospective opportunities with the College of Information Science and Electronic Engineer-
and listing some potential challenges. Finally, we ing, Zhejiang University, China. He received the Ph.D. degree in
provided a use case on how to obtain green- communication and information systems from the PLA Univer-
er 5G cellular networks and demonstrated the sity of Science and Technology, Nanjing, China, in 2002. From
September 2002 to December 2004, he was a postdoctoral
thrilling effectiveness of AI. We could boldly researcher at Zhejiang University. His research areas include
argue that AI empowered 5G cellular networks cognitive radio, wireless multi-hop networks (ad hoc, mesh,
will successfully enter the central stage of a digi- WSN, etc.), wireless multimedia network and green communi-
talized world. cations.

Xuan Zhou is a senior architect with the Service Provider Oper-

Acknowledgment ation Lab (SPO Lab) of Huawei Technologies. He received his
This article is supported by the National Postdoc- Ph.D. in communication and information systems from Zhejiang
toral Program for Innovative Talents of China (No. University, Hangzhou, China. From 2009 to 2014, he worked
as a system engineer at China Mobile Zhejiang Company. His
BX201600133), the Program for Zhejiang Lead- research efforts focus on innovative service and network man-
ing Team of Science and Technology Innovation agement in 5G, network function virtualization, and software-de-
(No. 2013TD20), the Zhejiang Provincial Technol- fined networks. He is also the architect of the world’s first 5G
ogy Plan of China (No. 2015C01075), the Nation- end-to-end network slicing demo, which was shown at Mobile
World Congress (MWC) 2016 in Barcelona, Spain.
al Natural Science Foundation of China (No.
61501510), and the Natural Science Foundation G uoru D ing [S’10, M’14, SM’16] received his B.S. degree
of Jiangsu Province (Grant No. BK20150717). (hons.) in electrical engineering from Xidian University, Xi’an,
China, in 2008, and his Ph.D. degree (hons.) in communications
References and information systems from the College of Communications
[1] C. X. Wang et al., “Cellular Architecture and Key Technolo- Engineering, Nanjing, China, in 2014. Since 2014, he has been
gies for 5G Wireless Communication Networks,” IEEE Com- an assistant professor with the College of Communications Engi-
mun. Mag., vol. 52, no. 2, Feb. 2014, pp. 122–30. neering. Since April 2015, he has been a postdoctoral research
[2] J. G. Andrews et al., “What Will 5G Be?” arXiv:1405.2957 [cs, associate at the National Mobile Communications Research Lab-
math], May 2014; available: oratory, Southeast University, Nanjing, China. His research inter-
[3] D. Soldani and A. Manzalini, “Horizon 2020 and Beyond: ests include cognitive radio networks, massive MIMO, machine
On the 5G Operating System for a True Digital Society,” learning, and big data analytics over wireless networks. He cur-
IEEE Vehic. Tech. Mag., vol. 10, no. 1, Mar. 2015, pp. 32–42. rently serves as a guest editor of the IEEE Journal on Selected

182 IEEE Wireless Communications • October 2017

Areas in Communications. He was a recipient of the Best Paper Z hongyao W ang received a B.Sc. in mathematics from the
Awards from EAI MLICOM 2016, IEEE VTC 2014-Fall, and IEEE School of Mathematical Sciences, Peking University, Beijing,
WCSP 2009. China in 2009. He joined the Deep Algorithm Department of
Alibaba Group, Hangzhou, China in 2013, where he currently
Y an C hen received her B.Sc. and Ph.D. degrees from Zhe- is a staff engineer. From 2010 to 2013, he worked in ASML
jiang University in 2004 and 2009, respectively. She was a (Brion) as a software engineer. He has focused on anti-fraud,
visiting researcher at the University of Science and Technology prediction, and recommendation in the past few years. He is
(HKUST) from 2008 to 2009. In 2009, she joined Huawei now responsible for user profile and ID-mapping, which are two
Technologies (Shanghai) Co., Ltd. She was the team leader key fundamental infrastructures inside Alibaba group.
and project manager of the internal project Green Radio Excel-
lence in Architecture and Technology (GREAT) from 2010 to H onggang Z hang is a professor with Zhejiang University,
2013, during which time she was also the project leader of China, and was the International Chair Professor of Excellence
the umbrella project Green Transmission Technologies (GTT) at the Université Européenne de Bretagne and Supélec, France
at the GreenTouch TM Consortium. Since 2013, she has been from 2012 to 2014. He is also an honorary visiting professor
the technical leader and project manager of the internal 5G with the University of York, UK. He served as the chair of
air interface design project. Her current research interests are the Technical Committee on Cognitive Networks of the IEEE
more toward future communication system design to efficiently Communications Society from 2011 to 2012. He is currently
support the multiplexing of different service scenarios with involved in research on green communications, serving as the
diversified requirements. series editor for the IEEE Communications Magazine series on
Green Communications and Computing Networks.

IEEE Wireless Communications • October 2017 183