Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
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
Flexible RA response
symbol Message 3
Packets with preambles
length Contention resolution
resource allocation
Time Packets
X
Location
update Data Core
message packet networks
Location
update
message
(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.
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
Granted or grant-free
RRM Granted Flexible bandwidth ü ü û ü UE-specific on-demand
Flexible symbol length
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-
Global controller
AI center
Data interface
... Traffic UE Resource ...
Pico and macro Shared signal Gateways Switches MMEs MiddleBoxes
BSs processing pool
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
Critic
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.