SMART, OPTIMAL, AND EXPLAINABLE ORCHESTRATION OF NETWORK SLICES IN 5G AND BEYOND NETWORKS

AI-Native Network Slicing for 6G Networks

Wen Wu, Conghao Zhou, Mushu Li, Huaqing Wu, Haibo Zhou, Ning Zhang, Xuemin (Sherman) Shen, and Weihua Zhuang

Abstract ity and hologram video streaming. To support

With the global rollout of fifth generation (5G) diversified services as established in 5G networks,
networks, it is necessary to look beyond 5G and network slicing is a potential approach to construct
envision 6G networks. 6G networks are expect- multiple logically isolated virtual networks (i.e., slic-
ed to have space-air-ground integrated networks, es) for different services on top of the common
advanced network virtualization, and ubiquitous physical network [4]. The QoS requirements of
intelligence. This article presents an artificial intel- different services can be guaranteed via cost-ef-
ligence (AI)-native network slicing architecture for fective slice management strategies ranging from
6G networks to enable the synergy of AI and net- preparation, planning, and operation phases in the
work slicing, thereby facilitating intelligent network network slicing life cycle.
management and supporting emerging AI services. Developing network slicing schemes faces many
AI-based solutions are first discussed across the challenges in 6G networks due to their unique fea-
network slicing life cycle to intelligently manage tures. First, managing slices over space, air, and
network slices (i.e., AI for slicing). Then network ground network segments in the SAGIN requires
slicing solutions are studied to support emerging AI judicious coordination of heterogeneous network
services by constructing AI instances and perform- segments. Moreover, 6G networks need to sup-
ing efficient resource management (i.e., slicing for port a variety of new services while satisfying their
AI). Finally, a case study is presented, followed by different and stringent QoS requirements, which
a discussion of open research issues that are essen- further complicates slice management. Hence, it
tial for AI-native network slicing in 6G networks. is paramount to develop intelligent slice manage-
ment solutions in 6G networks. Second, fueled
Introduction by powerful computing capability and advanced
Compared to existing wireless networking includ- AI techniques, ubiquitous intelligence is fostering
ing the fifth generation (5G), 6G is more than abundant AI services with new QoS requirements,
an improvement of key performance indicator such as data quality, inference accuracy, and train-
(KPI) requirements, such as increased data rates, ing latency. Hence, it is necessary to construct cus-
enhanced network capacity, and low latency. 6G tomized network slices to support the emerging AI
networks are envisioned to have the following services in 6G networks.
unique features. First, space networks, such as In this article, we propose an AI-native network
low Earth orbit (LEO) satellites, air networks (e.g., slicing architecture for 6G networks to facilitate
unmanned aerial vehicles, UAVs), and ground intelligent network management while support-
networks, such as cellular base stations (BSs), are ing emerging AI services. AI-native means that, as
integrated into a space-air-ground integrated net- a built-in component in the network slicing archi-
work (SAGIN) to provide global coverage and tecture, AI exists not only in the software-defined
on-demand services [1]. Second, resource virtual- networking (SDN) controller for managing network
ization using network slicing techniques and end slices, but also in network slices as services for end
user virtualization using digital twin techniques can users. Hence, the synergy of AI and network slicing
facilitate advanced network virtualization to pro- in the proposed architecture is two-fold. On one
vide flexible network management [2, 3].1 Third, hand, AI techniques can be applied to manage net-
intelligence penetrates every corner of networks, work slices, namely AI for slicing. The network slic-
ranging from end users through the network edge ing life cycle including preparation, planning, and
to the remote cloud, which results in ubiquitous operation phases is introduced, along with specify-
intelligence. A number of network nodes are ing AI-based solutions for each phase. In addition,
endowed with built-in artificial intelligence (AI) the detailed procedure of information exchange
functionalities, thereby not only facilitating intelli- among end users, access points, and the SDN con-
1 End user virtualization is to
gent network management but also fostering AI troller is presented. On the other hand, network
virtualize end users in net- services (e.g., deep neural network based applica- slicing can be applied to construct customized net-
work operation and manage- tions). Hence, 6G networks are expected to create work slices for various AI services, namely slicing
ment by characterizing users’ a new wireless networking ecosystem that brings for AI. Potential approaches such as AI instance
behaviours and status (e.g., societal and economic benefits. construction and efficient resource management
service demands and QoS
satisfaction), which can be The 6G networks will support a diverse set of for AI services are introduced.
achieved using digital twin services with different quality of service (QoS) The remainder of this article is organized as
concepts. requirements, such as multisensory extended real- follows. Expected features of 6G networks are

Digital Object Identifier: Wen Wu is with the Peng Cheng Laboratory, China; Conghao Zhou, Mushu Li, Huaqing Wu, Xuemin Shen, and Weihua Zhuang are with the
10.1109/MWC.001.2100338 University of Waterloo, Canada; Haibo Zhou is with Nanjing University, China; Ning Zhang is with the University of Windsor, Canada.

96 1536-1284/22/$25.00 © 2022 IEEE IEEE Wireless Communications • February 2022

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.
 discussed, and then the AI-native network slicing virtual reality (VR) and hologram video streaming With caching capabil-
architecture is proposed. The basic ideas of AI for applications require a high data rate; for example, ity, a large amount of
slicing and slicing for AI are presented, respective- the uplink data rate of mobile VR is up to 5 Gb/s. data can be stored in
ly. A case study is presented. The research direc- Other applications may require ultra-high reliabili- the network. In addi-
tions are identified, followed by the conclusion. ty, such as autonomous driving, industrial control
systems, and robot/UAV swarm; for example, the tion, with the develop-
AI-Native Network Slicing for 6G Networks required reliability of autonomous driving is up to ment of AI techniques,
99.999 percent [10] edge computing, and
Network Slicing • Ubiquitous intelligence: With caching capabili- device computing,
Network slicing is an emerging technology to ty, a large amount of data can be stored in the intelligence is pushed
support diversified applications in a cost-effec- network. In addition, with the development of
tive manner [4, 5]. The concept of network slicing AI techniques, edge computing, and device from the remote cloud
can be traced back to the late 1980s [6]. Now- computing, intelligence is pushed from the to the network edge
adays, network slicing is a key technology in 5G remote cloud to the network edge and end and end users.
networks, supported by network function virtual- users. As such, AI will be integrated into 6G
ization (NFV) and SDN techniques. Specifically, networks for intelligent network management
NFV enables virtualized resources and network by directly learning from extensive data in the
functions for flexible resource management, while network. Moreover, ubiquitous intelligence
SDN facilitates centralized network management will foster a number of AI services in which AI
for network optimization. In 5G networks, net- is provided as services.
work slicing has been defined in the 3rd Gener-
ation Partnership Project (3GPP) Release 15 [7]. AI-Native Network Slicing Architecture
Moreover, in coming 6G networks, network slic- These features impose new challenges on devel-
ing will continue to evolve and play an increasing- oping network slicing schemes for 6G networks.
ly important role. First, the SAGIN not only increases the number of
The basic idea of network slicing is to create integrated network segments but also introduces
multiple logically isolated network slices on top extra dynamics on network resource availability
of the common physical infrastructure, which can due to satellite mobility and UAV manoeuverabil-
achieve flexible and adaptive network manage- ity. Network slicing schemes should accommo-
ment. Its benefits are three-fold: date the large-scale SAGIN while taking dynamic
1. Multi-tenancy: Multiple virtual networks can resource availability into account. Moreover, sup-
share the common physical infrastructure, thus porting diversified services with stringent QoS
reducing capital expenditures in the network requirements further complicates network slicing
deployment. scheme design. Second, ubiquitous intelligence
2. Service isolation: Multiple slices are constructed facilitates many emerging AI services that will be
for different services via judicious resource man- prevalent in 6G networks. Different from con-
agement, such that service level agreements of ventional services, facilitating AI services requires
different slices can be effectively guaranteed. multiple steps, including collecting high-quality
3. Flexibility: Network slicing can support flexible data samples, training satisfactory AI models, and
network management, as slices can be created, performing low-latency model inference, which
modified, or deleted on demand. should meet diverse QoS requirements. How to
satisfy such diverse QoS requirements for AI ser-
Features of 6G Networks vices remains a challenging issue.
From 5G to 6G, it is in general expected that To address the above challenges, an AI-native
KPI requirements will be increased by at least an network slicing architecture for 6G networks is
order of magnitude. According to a recent white presented. As shown in Fig. 1, the architecture
paper [8], the KPI requirements of 6G networks aims to integrate SAGIN and ubiquitous intel-
include 1 Tb/s peak data rate, 20–100 Gb/s user ligence and support diverse services with strin-
experienced data rate, 0.1 ms end-to-end latency, gent QoS requirements. Compared to network
10 million devices/km 2, and near 100 percent slicing for 5G networks, the proposed architec-
coverage. Such KPI requirements demand several ture has two new characteristics. First, AI is inte-
candidate technologies, such as THz communi- grated into SDN controllers to realize intelligent
cations and AI [6]. The 3GPP working group will network slicing such that a number of network
discuss 6G candidate techniques by the end of slices with stringent QoS requirements can be
2026, and the first 6G standard is expected to managed efficiently and cost-effectively via AI
debut by 2030. techniques, which is referred to as AI for slicing.
Distinguished from 5G networks, 6G networks Second, emerging AI services are supported by
have several features: network slicing. In addition to network slices for
• SAGIN: While current ground networks provide conventional services, new network slices are
good coverage in highly populated areas, 6G constructed for AI services on top of the com-
needs to provide universal coverage, including mon physical infrastructure, which is referred to
in rural, remote, and sparsely populated areas. as slicing for AI.
To achieve this goal, 6G will exploit the alti- Two types of SDN controllers are deployed
tude dimension. Space, air, and ground net- in the proposed architecture. One is the central-
work segments are integrated into the SAGIN ized SDN controller located at the cloud, which
[1, 9], which can provide global coverage, facil- is to manage network slices. The other is the local
itate on-demand services, and support high-rate SDN controller located at access points, which is
low-delay services; to schedule resources to end users within each
• Diversified services: Many services have stringent network slice. The SDN controller in the follow-
QoS requirements in different dimensions. Mobile ing refers to the centralized SDN controller unless

IEEE Wireless Communications • February 2022 97

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.
 WU_1.23.pdf 1 3/19/22 1:04 AM

the corresponding procedure of information

exchange in AI for slicing is discussed.
Centralized SDN
Controller
network slIcIng lIfe cycle
High Reliability
The network slicing life cycle consists of three
phases: preparation, planning, and operation, as
Massive Connectivity shown in Fig. 2. The centralized SDN controller is
in charge of the preparation and planning phases,
Low Delay while the operation phase is coordinated by local
SDN controllers.
AI Service High Accuracy Preparation Phase: This phase is to construct and
configure network slices based on service require-
C
ments, data traffic, user information, and virtual net-
work resource availability. To achieve the goal, the
M

SDN controller conducts the following tasks:

CM

MY

• Service requirement extraction: This task is

CY

Space and Air Network

CMY

to classify services by extracting their QoS

K

requirements, such as service delay, service

Core Network priority, throughput, and reliability. 3GPP has
standardized specific service/slice type values
for classified services, such as enhanced mobile
broadband, ultra-reliable low-latency communi-
Ground Network cations, and massive machine-type communica-
tions services [7].
User with AI Service AI Function Switch Local SDN Controller • Network resource and function virtualization:
Network resources, such as communication,
Satellite UAV Access Point computing, and caching resources, are pooled
into virtualized resource blocks via advanced
FIGURE 1. An illustration of the AI-native network slicing architecture for 6G resource virtualization techniques. Similarly,
networks. network functions, such as firewall, network
WU_2.23.pdf 1 3/19/22 1:05 AM
name translation, and domain name system,
are separated from dedicated hardware net-
Communication Computing Caching work functions into virtualized network func-
Resource Resource Resource
tions (VNFs). Through virtualization, the SDN
controller can flexibly manage network resourc-
Delay Throughput Reliability Accuracy ... es and functions.
Virtualized Network Resource Pool
Once these tasks are completed, the SDN con-
Service Requirement Extraction
troller can construct network slices for each
admitted slice request.
User Mobility
Pattern
Firewall Planning Phase: This phase aims to reserve net-
Channel Condition Service Demand
work resources to slices for service provisioning.
Network Name Domain Name
Stochastics Pattern Translation System The planning phase operates over a long timescale.
Centralized SDN Time is partitioned into multiple planning periods
Stochastic Network Controller Virtualized Network
C

(windows) for each slice. The duration of each

M

Information Function
Network Slicing planning window depends on service demand and
CM

MY
Feedback
CY

network dynamics, whose value ranges from sev-

eral minutes to several hours. To achieve the goal,
CMY

Resource VNF Resource

Reservation Placement Orchestration
RAT Selection
the following two steps are conducted in the plan-
ning phase:
Resource Resource Block • Service and network information collection: Ben-
Communication
Resource User 1 efiting from the global control functionality of
Slice 1 Planning
Computing Decision the SDN controller, extensive network informa-
User 2
Slice 2 Resource tion can be collected from underlying physical
Caching
User 3 networks, such as service demands, stochastic
Slice N
Resource
Time channel conditions, and user mobility patterns.
Planning Operation The collected information is utilized for the fol-
lowing resource reservation decision making.
• Resource reservation: At the beginning of each
FIGURE 2. An illustrative example of the network slicing life cycle, including planning window, the SDN controller adjusts
the preparation, planning, and operation phases. the amount of reserved network resources for
each slice based on the monitored slice per-
formance. The reserved virtualized network
otherwise stated. In the following, we illustrate resources of each slice are mapped to the phys-
the basic ideas of AI for slicing and slicing for AI, ical network. At the end of each planning win-
respectively. dow, some system information is fed back to
the SDN controller, such as resource utilization,
AI for slIcIng system performance, and service level agree-
In this section, we introduce the network slic- ment satisfaction. Based on the feedback infor-
ing life cycle in three phases and then investi- mation, the SDN controller can adjust resource
gate potential AI solutions for each phase. Next, reservation decisions to accommodate dynam-

98 IEEE Wireless Communications • February 2022

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.
 WU_3.23.pdf 1 3/19/22 1:06 AM

ic network environments while guaranteeing

QoS requirements.
Operation Phase: This phase is to schedule the Preparation
service of a slice using the reserved resources for Historical
Service Demand
subscribed end users. The operation phase works
over a much shorter timescale (e.g., 100 ms) than Slice
Admission
that of the planning phase. Specifically, under the
coordination of the centralized SDN controller, Resource ...
C

Availability ...
M

local SDN controllers allocate network resources Y

Planning ...
to end users in each slice according to their real- ...
CM

Resource
MY

time data traffic. The operation decisions include

CY

Reservation
AI
CMY

selecting radio access technology (RAT), deter- K

Planning
mining user association with specific radio access Decision
points, deciding proper protocol and associated User Mobility,
Channel Condition
parameters, and orchestrating resources among Physical
end users. Network Operation
Roles of AI in Network Slicing: Although net- Resource
Orchestration
work slicing can facilitate service provisioning,
managing a number of network slices incurs sig- FIGURE 3. The considered AI-based network slicing solution in which AI plays
nificant network management cost, especially in different roles in the preparation, planning, and operation phases.
6G networks. As shown in Fig. 3, AI-based network
slicing is a potential solution in which AI plays dif-
ferent roles in different network slicing phases. didate RATs for each end user. Due to user
AI for preparation: In the preparation phase, AI mobility, user-perceived service performance
needs to perform two tasks. of an RAT is stochastic. This problem can be
1. Service demand prediction: Based on historical addressed by multi-armed bandit methods (e.g.,
data, service demand can be predicted via AI contextual bandit).
techniques, such as recurrent neural networks.
Prior studies show that the service demand and procedure of InformAtIon eXchAnge
resource usage of a slice can be accurately pre- The AI for slicing procedure involves the informa-
dicted [11]. The prediction results can be uti- tion exchange among end users, access points,
lized for decision making in the planning phase. and the SDN controller. The procedure is illustrat-
2. Slice admission: The SDN controller admits ed in Fig. 4 with the following steps:
slices to maximize network resource utilization 1. Access points collect user-level stchastic informa-
considering resource availability and service tion, such as end users’ service demand patterns,
demands. As the slice admission decision is mobility patterns, and stochastic channel conditions.
binary, this problem is deemed as an integer 2. Access points translate the user-level infor-
optimization problem. In large-scale networks mation into desired service-level information.
with complex resource availability distribution, For example, user density information can be
conventional optimization solutions become obtained from processing user location infor-
complicated and intractable, while AI-based mation, and AI techniques can be used for such
solutions have potential. data abstraction, fusion, and analysis; . RNN
AI for planning: In the planning phase, AI can can be applied for service demand prediction.
perform two tasks: 3. The processed service-level information is deliv-
1. VNF placement: The SDN controller deploys ered to the SDN controller.
VNFs to support services in the network. The 4. The SDN controller runs AI-based planning
resources allocated for VNFs should be dynam- algorithms to make decisions based on the col-
ically adjusted for time-varying service demands lected service-level information.
to guarantee service delay requirements. Deep 5. The determined planning decisions are sent
learning methods can be applied to enhance back to all access points.
resource utilization in dynamic network envi- 6. Access points enforce the received planning
ronments. decisions (e.g., reserving network resources for
2. Resource reservation: The SDN controller corresponding slices).
reserves resources for different slices based 7. End users in service report their real-time infor-
on their service demands. Since data traffic mation to their associated access points, such
loads are time-varying, the resource reserva- as real-time service demands, channel condi-
tion should be adaptive to dynamic real-time tions, and task data sizes.
demands, which can be addressed via rein- 8. Access points run the AI-based operation algo-
forcement learning (RL) methods, such as deep rithm to allocate resources for end users based
deterministic policy gradient (DDPG). on real-time user-level information.
AI for operation: Two exemplary operation 9. Service requests from end users are support-
tasks are: ed with the allocated network resources. For
1. Resource orchestration: The reserved resources example, computation tasks can be offloaded
of a slice are allocated to end users. The deci- to access points using communication resource
sions are determined based on real-time user and then processed using computing resourc-
mobility, service demands, and so on. To effi- es. For each operation slot within a planning
ciently utilize resources, RL methods can be window, Steps 7–9 are repeated;
applied for dynamic resource orchestration; 10. Access points monitor slice performance in the
2. RAT selection: To maximize system utility, an network given the enforced planning decisions
optimal RAT is selected among multiple can- by measuring end users’ satisfaction rates across

IEEE Wireless Communications • February 2022 99

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.
 WU_4.23.pdf 1 3/19/22 1:12 AM

example, objective detection services can be

End Users Access Points SDN Controller implemented via ResNet32, Inception-v3, AlexNet,
or VGG16 algorithms. Hence, the primary issue of
1. User-Level Stochastic Information supporting an AI service is to determine an appro-
2. AI-Based Information
priate implementation option in the network.
Processing We introduce the concept of AI instance
3. Service-Level Information to address the issue, as shown in Fig. 5. An AI
instance of an AI service represents an implemen-
4. AI-Based Planning
Decision Making
tation option for an AI service. The basic idea is
5. Planning Decision
to construct multiple candidate implementation
options and then select an appropriate one based
C

M Planning
6. Planning Decision
Y

Enforcement
Window on network environments. The procedure of the
conceptual AI instance management framework
CM

MY

7. User-Level Real-Time Information

CY
consists of two steps:
1. AI instance construction: The network opera-
CMY

8. AI-Based Operation
K
Operation
Decision Making
Slot tor constructs multiple candidate AI instances
9. Data for each AI service based on available network
resources and service requirements. An AI
Repeat 7-9 During Each Operation Slot instance may include:
10. Slice Performance –The AI algorithm, which specifies the imple-
Measurement and Monitoring mentation algorithm and the corresponding
11. Slice Performance neural network architecture
12. Adjust AI-Based Planning –The training manner of the AI algorithm (e.g.,
Policy centralized or distributed training)
–The amount of required network resources
FIGURE 4. Procedure of information exchange in AI for slicing.
WU_5.23.pdf 1 3/19/22 1:10 AM
2. AI instance selection: In this step, the AI ser-
vice provider selects an appropriate AI instance
among candidate AI instances based on user
... service preference (e.g., privacy preservation
User
preference). If an AI instance is selected, the AI
Extract service will be executed using the AI algorithm
User Information and the corresponding required amount of net-
work resources by the AI instance. In summary,
... AI Service
the idea of AI instance provides flexibility for AI
Provider
service management.
C

AI Algorithm Select
M

The network operator can dynamically manage

Y

AI Instance
Training Manner
CM

AI services via creating, modifying, and deleting

MY

Computing Resource
CY

CMY
AI
Instance
AI
Instance
AI
Instance
AI
Instance
AI
Instance
AI
Instance
candidate AI instances on demand. Note that AI
Communication Resource ... ... ... ...
K

instances are not limited to neural networks, and

Caching Resource
they can include other types of AI algorithms such
Construct
as support vector machines and decision trees.
AI Instance

Network
resource mAnAgement In AI servIce lIfe cycle
Operator
Virtualized Network Resource Pool
Running AI services includes three stages: data
collection, model training, and model inference
FIGURE 5. The conceptual AI instance management framework for AI services. (i.e., the AI service life cycle) [12, 13]. Specifi-
cally, data collection is to collect data via com-
all operation slots within a planning window. munication links, and the collected data can be
11. Access points report network performance to stored in network edge servers. Based on the col-
the SDN controller. lected data, an AI model can be trained in the
12. The SDN controller makes the planning deci- model training stage. The model training can be
sion for the next planning window and adjusts implemented in either a centralized or distributed
the planning policy based on the feedback manner. For example, multiple devices can work
information. collaboratively to train a global model via feder-
Note that in the preceding procedure, Steps 1–6 ated learning. Next, well-trained AI models are
and Steps 10–12 are in the network planning deployed to execute specific computation tasks;
phase, and Steps 7–9 are in the operation phase. this is referred to as model inference. The model
inference can be performed in multiple ways.
slIcIng for AI For example, device-edge collaborative inference
Slicing for AI utilizes network slicing to support AI approaches can allocate and process computa-
services while satisfying QoS requirements. Poten- tion tasks at different network nodes to achieve
tial solutions include constructing and selecting AI low inference latency.
instances and efficient resource management in The performance of an AI service depends
the AI service life cycle. on all three stages of the AI service life cycle. For
example, model inference accuracy depends on
AI InstAnce multiple factors, such as the quality of the collect-
There are diversified implementation options for ed data, the number of training iterations, and
supporting AI services. An AI service can be imple- the approach of model inference. Meanwhile, all
mented via different kinds of algorithms, training three stages consume multi-dimensional network
manners, and network resource allocation. For resources. As a result, to optimize the performance

100 IEEE Wireless Communications • February 2022

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.
 of AI services, network resources should be joint- 3000
The cumulative system
ly allocated for these three stages. The reserved DDPG Solution cost within one day
network resources in AI slices should be further is presented. It can
2500
allocated to these three stages to satisfy their corre- be observed that the

Overall System Cost

sponding QoS requirements.
2000 DDPG-based solution
Case Study can reduce the cumu-
In this section, a case study is provided on AI-as- 1500 lative overall system
sisted resource reservation, aimed at reducing cost within one day by
long-term overall system cost. 1000 around 15 percent as
Considered Scenario compared to the myo-
We consider an air-ground integrated network for 500
0 1000 2000 3000 4000
pic solution.
providing autonomous driving services to vehicles Training Episodes
traversing a highway segment. For the considered (a)
highway segment with a length of 2 km, two BSs 800
are uniformly deployed along the highway with 700
Myopic Solution
DDPG Solution
a separation distance of 1 km, and one UAV is 15%

Cumulative System Cost

600
deployed in the centre hovering at a height of
100 m. When an autonomous vehicle is driving 500
on the highway, extensive computation-intensive
400
tasks must be processed. For prompt task pro-
cessing, all access points are equipped with edge 300
computing servers, and vehicles can associate 200
with the nearest access point and upload their
computation tasks. We consider a delay-sensitive 100
autonomous driving service, object detection, 0
whose delay requirement is 100 ms for road safe- 0 5 10 15 20 25
ty [14]. The data size and computation intensity Time (Hour)
(b)
of a task are set to 0.6 Mbit and 6  108 cycles,
respectively. The service delay is characterized by FIGURE 6. Performance evaluation of the pro-
the queuing theory since task arrivals are assumed posed DDPG-based resource reservation
to follow a Poisson process with rate l = 1 pack- solution: a) convergence performance; b)
et/s. To guarantee the service delay requirement, cumulative system cost within one day.
a network slice is constructed in which spectrum
and computing resources are reserved.
Resource reservation decisions are made
Simulation Results
to minimize the overall system cost considering We evaluate the performance of the proposed
vehicle traffic dynamics. The overall system cost is DDPG-based solution based on real-world high-
defined as C = STt=1(wr Crt + ws Cst + wdCdt ), which is way vehicle traffic flow trace collected by Alber-
a weighted summation of three cost components ta Transportation.2 As shown in Fig. 6a, we first
across all T planning windows. present the convergence performance of the pro-
1. Resource reservation cost Crt accounts for the posed DDPG-based solution. A five-point mov-
amount of the reserved spectrum and comput- ing average is applied to process raw simulation
ing resources at BSs at planning window t. The points to highlight the convergence trend (i.e.,
spectrum resource is allocated in a unit of sub- red curve). It can be seen that the DDPG-based
carrier of 5 MHz, and the computing resource resource reservation solution has converged after
is allocated in a unit of virtual machine (VM) 4000 training episodes.
instance with a processing rate of 10  10 9 Next, as shown in Fig. 6b, the cumulative system
cycles/s. cost within one day is presented. It can be observed
2. Slice reconfiguration cost Cst accounts for the that the DDPG-based solution can reduce the
difference between two consecutive resource cumulative overall system cost within one day by
reservation decisions [15]. around 15 percent as compared to the myopic solu-
3. Delay requirement violation penalty Cdt refers tion. The reason is that the proposed DDPG-based
to the penalty once service delay exceeds the solution is able to minimize the long-term overall
delay requirement. These weight parameters are system cost, while the myopic solution minimizes
set to wr = 1, ws = 20, and wd = 200, respectively. the short-term system cost, which incurs prohibitive
The planning window size is set to one hour. slice reconfiguration cost due to frequent adjust-
We propose a DDPG-based solution to mini- ment of network resource reservation in highly
mize the overall system cost [15]. In this solution, dynamic vehicular networks. The simulation results
both actor and critic networks are fully connect- show that the proposed AI-based resource reserva-
ed neural networks with four layers, the numbers tion solution can achieve a low system cost.
of neurons in two hidden layers are 128 and 64,
respectively, and their learning rates are set to 2  Open Research Issues
10–4 and 2  10–3, respectively. For performance In the following, we discuss some open research
comparison, we adopt an optimization-based solu- issues pertaining to AI-native network slicing.
tion, named myopic resource reservation, in which
network resources are reserved to minimize the Joint Design of Network Planning and Operation 2 Alberta Transportation;
resource reservation cost at each planning window Planning and operation are performed and cou- http://www.transportation.
while satisfying the delay requirement. pled in two different timescales. Specifically, the alberta.ca/mapping/.

IEEE Wireless Communications • February 2022 101

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.
 The slicing for AI is to planning phase is performed in a long timescale [4] X. Shen et al., “AI-Assisted Network-Slicing Based Next-Gen-
eration Wireless Networks,” IEEE Open J. Vehic. Tech., vol. 1,
construct customized (e.g., minute level) to reserve resources for dif- no. 1, Jan. 2020, pp. 45–66.
network slices to better ferent slices based on service demands, while the [5] W. Zhuang et al., “SDN/NFV-Empowered Future IoV with
accommodate various operation phase is performed in a short timescale Enhanced Communication, Computing, and Caching,” Proc.
(e.g., sub-second level) to allocate the reserved IEEE, vol. 108, no. 2, Feb. 2020, pp. 274–91.
emerging AI services. [6] X. You et al., “Towards 6G Wireless Communication Net-
resources to on-demand users within each slice. works: Vision, Enabling Technologies, and New Paradigm
To accelerate the pace Achieving the optimal network slicing perfor- Shifts,” Sci. China Info. Sci., vol. 64, no. 1, Jan. 2021, pp.
of AI-native network mance requires a joint optimization design of 1–74.
planning and operation. [7] A. Kaloxylos, “A Survey and an Analysis of Network Slicing in
slicing architecture 5G Networks,” IEEE Commun. Stds. Mag., vol. 2, no. 1, Mar.
development, extensive 2018, pp. 60–65.
research efforts are
Data Management Framework [8] N. Rajatheva et al., “White Paper on Broadband Connec-
The cornerstone of AI-native network slicing is tivity in 6G,” arXiv:2004.14247, 2020; https://arxiv.org/
required, such as in abundant data that can be used for AI model abs/2004.14247.
[9] C. Zhou et al., “Deep Reinforcement Learning for Delay-Ori-
the identified research training. In 6G networks, data is widely distributed ented IoT Task Scheduling in Space-Air-Ground Integrated
directions. in the network. Due to limited communication Network,” IEEE Trans. Wireless Commun., vol. 20, no. 2, Feb.
resources, the cost of collecting a large amount 2021, pp. 911–25.
of data cannot be neglected. In addition, the col- [10] C. Campolo et al., “5G Network Slicing for Vehicle-to-Ev-
erything Services,” IEEE Wireless Commun., vol. 24, no. 6,
lected data needs to be processed to mine valu- Dec. 2017, pp. 38–45.
able information for network management. For [11] C. Gutterman et al., “RAN Resource Usage Prediction for a
example, abundant historical behaviour data from 5G Slice Broker,” Proc. ACM MobiHoc, Catania, Italy, 2019.
individual users can be analyzed to predict spa- [12] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning,
MIT Press, 2016.
tio-temporal service demand distributions. Hence, [13] M. Li et al., “Slicing-Based Artificial Intelligence Service
establishing a data management framework to Provisioning on the Network Edge: Balancing AI Service
collect and analyze data is necessary. Performance and Resource Consumption of Data Manage-
ment,” IEEE Vehic. Tech. Mag., vol. 16, no. 4, Dec. 2021,
Prediction-Empowered Network Slicing pp. 16–26..
[14] S.-C. Lin et al., “The Architectural Implications of Autono-
With the development of advanced AI technol- mous Driving: Constraints and Acceleration,” Proc. ASPLOS,
ogies, the data traffic in the network can be pre- 2018, pp. 751–66.
dicted. How to effectively leverage the power [15] W. Wu et al., “Dynamic RAN Slicing for Service-Oriented
Vehicular Networks via Constrained Learning,” IEEE JSAC,
of prediction for network slicing is an interesting vol. 39, no. 7, 2021, pp. 2076–89.
topic. Since the prediction is imperfect, the pre-
diction error may degrade the performance of Biographies
network slicing. How to evaluate the impact of Wen Wu [S’13, M’20] ([email protected]) earned his Ph.D.
prediction errors on system performance and to degree in electrical and computer engineering from the Univer-
sity of Waterloo, Ontario, Canada, in 2019. He received his B.E.
develop corresponding solutions are important degree in information engineering from South China University
research issues. of Technology, Guangzhou, China,and his M.E. degree in elec-
trical engineering from University of Science and Technology
Conclusion of China, Hefei, in 2012 and 2015, respectively. He worked
as a postdoctoral fellow with the Department of Electrical and
In this article, we have proposed the AI-native net- Computer Engineering, University of Waterloo. He is currently
work slicing architecture to facilitate intelligent an associate researcher at Peng Cheng Laboratory, Shenzhen,
network management and support AI services in China. His research interests include 6G networks, pervasive
6G networks. The architecture aims to enable the network intelligence, and network virtualization.
synergy of AI and network slicing. AI for slicing Conghao Zhou [S’19] ([email protected]) received his
helps reduce network management complexity, B.E. degree from Northeastern University, Shenyang, China, in
while adapting to dynamic network environments 2017 and his M.S. degree from the University of Illinois at Chica-
by exploiting the capability of AI in network go in 2018. He is currently working toward a Ph.D. degree with
the Department of Electrical and Computer Engineering, Univer-
slicing. Slicing for AI is to construct customized sity of Waterloo. His research interests include space-air-ground
network slices to better accommodate various integrated networks and machine learning in wireless networks.
emerging AI services. To accelerate the pace of
AI-native network slicing architecture develop- Mushu Li [S’17, M’21] ([email protected]) is a postdoctoral
fellow with the Department of Electrical and Computer Engi-
ment, extensive research efforts are required, neering, University of Waterloo, where she also received her
such as in the identified research directions. Ph.D. degree in electrical engineering in 2021. She received
her M.A.Sc. degree from Ryerson University in 2017. She was
Acknowledgments the recipient of a Natural Science and Engineering Research
Council of Canada Graduate Scholarship in 2018 and an Ontar-
This work was financially supported by the Nat- io Graduate Scholarship in 2015 and 2016, respectively. Her
ural Sciences and Engineering Research Council research interests include machine learning in wireless networks
(NSERC) of Canada. The authors would like to and network slicing.
thank Jie Gao, Qihao Li, and Kaige Qu for many
Huaqing Wu [S’15, M’21] ([email protected]) received
valuable discussions and suggestions throughout her Ph.D. degree from the University of Waterloo in 2021.
the work. She received her B.E. and M.E. degrees from Beijing University
of Posts and Telecommunications, China, in 2014 and 2017,
References respectively. She received the prestigious Natural Sciences and
Engineering Research Council of Canada (NSERC) Postdoc-
[1] N. Zhang et al., “Software Defined Space-Air-Ground Inte-
grated Vehicular Networks: Challenges and Solutions,” IEEE toral Fellowship Award in 2021. She is currently a postdoctor-
Commun. Mag., vol. 55, no. 7, July 2017, pp. 101–09. al research fellow at McMaster University, Ontario, Canada.
[2] X. Shen et al., “Holistic Network Virtualization and Per- Her current research interests include vehicular networks with
vasive Network Intelligence for 6G,” IEEE Commun. Sur- emphasis on edge caching, wireless resource management,
veys & Tutorials, to be published, 2022. DOI: 10.1109/ space-air-ground integrated networks, and application of artifi-
COMST.2021.3135829. cial intelligence for wireless networks.
[3] R. Minerva, G. M. Lee, and N. Crespi, “Digital Twin in the
IoT Context: A Survey on Technical Features, Scenarios, Haibo Zhou [M’14, SM’18] ([email protected]) received
and Architectural Models,” Proc. IEEE, vol. 108, no. 10, Oct. his Ph.D. degree in information and communication engineering
2020, pp. 1785–1824. from Shanghai Jiao Tong University, China, in 2014. He is cur-

102 IEEE Wireless Communications • February 2022

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.
 rently an associate professor with the School of Electronic Sci- Xuemin (Sherman) Shen [M’97, SM’02, F’09] (sshen@uwater-
ence and Engineering, Nanjing University, China. He was named loo.ca) is currently a university professor with the Department
a 2020 cross-field highly cited researcher. He was a recipient of of Electrical and Computer Engineering, University of Water-
the 2019 IEEE Communications Society Asia-Pacific Outstanding loo. His research focuses on network resource management,
Young Researcher Award. He is currently an Associate Editor wireless network security, social networks, 5G and beyond,
of the IEEE Internet of Things Journal, IEEE Network, and IEEE and vehicular ad hoc networks. He is a Canadian Academy of
Wireless Communications Letters. His research interests include Engineering Fellow, a Royal Society of Canada Fellow, and a
resource management in VANETs, 5G/B5G wireless networks, Chinese Academy of Engineering Foreign Fellow. He received
and SAGINs. the R.A. Fessenden Award in 2019 from IEEE, Canada; the
James Evans Avant Garde Award in 2018 from the IEEE Vehic-
Ning Zhang [M’15, SM’18] ([email protected]) is cur- ular Technology Society; and the Joseph LoCicero Award in
rently an associate professor in the Department of Electrical and 2015 and Education Award in 2017 from the IEEE Communi-
Computer Engineering at the University of Windsor, Canada. cations Society.
He received his Ph.D. degree in electrical and computer engi-
neering from the University of Waterloo in 2015. His research Weihua Zhuang [M’93, SM’01, F’08] (wzhuang@uwaterloo.
interests include connected vehicles, mobile edge computing, ca) has been with the Department of Electrical and Computer
wireless networking, and machine learning. He is a Highly Cited Engineering, University of Waterloo since 1993, where she is cur-
Researcher. He serves as an Associate Editor of the IEEE Internet rently a professor and a Tier I Canada Research Chair in Wireless
of Things Journal, IEEE Transactions on Cognitive Communica- Communication Networks. She is a Fellow of the Royal Society
tions and Networking, and the IEEE Systems Journal; and has of Canada, the Canadian Academy of Engineering, and the Engi-
served as a Guest Editor for several international publications, neering Institute of Canada. She is also an elected member of the
such as IEEE Wireless Communications and IEEE Transactions on Board of Governors and VP-Publications of the IEEE Vehicular
Industrial Informatics. Technology Society.

IEEE Wireless Communications • February 2022 103

Authorized licensed use limited to: Amrita Vishwa Vidyapeetham Chennai Campus. Downloaded on September 18,2022 at 22:41:38 UTC from IEEE Xplore. Restrictions apply.