See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/351685332

AI-Native Network Slicing for 6G Networks

Preprint · May 2021

DOI: 10.48550/arXiv.2105.08576

CITATIONS READS
0 1,882

9 authors, including:

Wen Wu Conghao Zhou

Peng Cheng Laboratory University of Waterloo
127 PUBLICATIONS 3,528 CITATIONS 69 PUBLICATIONS 2,639 CITATIONS

SEE PROFILE SEE PROFILE

Mushu Li Huaqing Wu
Toronto Metropolitan University The University of Calgary
79 PUBLICATIONS 2,206 CITATIONS 80 PUBLICATIONS 2,338 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Wen Wu on 05 November 2021.

The user has requested enhancement of the downloaded file.

 1

AI-Native Network Slicing for 6G Networks

Wen Wu, Member, IEEE, Conghao Zhou, Student Member, IEEE, Mushu Li, Student Member, IEEE,
Huaqing Wu, Member, IEEE, Haibo Zhou, Senior Member, IEEE, Ning Zhang, Senior Member, IEEE,
Xuemin (Sherman) Shen, Fellow, IEEE, and Weihua Zhuang, Fellow, IEEE

Abstract—With the global roll-out of the fifth generation e.g., deep neural network based applications. Hence, 6G
(5G) networks, it is necessary to look beyond 5G and en- networks are expected to create a new wireless networking
vision the 6G networks. The 6G networks are expected to ecosystem that brings societal and economic benefits.
have space-air-ground integrated networks, advanced network
virtualization, and ubiquitous intelligence. This article presents The 6G networks will support a diverse set of services
an artificial intelligence (AI)-native network slicing architecture with different quality of service (QoS) requirements, such as
for 6G networks to enable the synergy of AI and network multisensory extended reality and hologram video streaming.
slicing, thereby facilitating intelligent network management and To support diversified services as established in 5G networks,
supporting emerging AI services. AI-based solutions are first network slicing is a potential approach to construct multiple
discussed across network slicing lifecycle to intelligently manage
network slices, i.e., AI for slicing. Then, network slicing solutions logically-isolated virtual networks (i.e., slices) for different
are studied to support emerging AI services by constructing AI services on top of the common physical network [4]. The
instances and performing efficient resource management, i.e., QoS requirements of different services can be guaranteed
slicing for AI. Finally, a case study is presented, followed by a via cost-effective slice management strategies ranging from
discussion of open research issues that are essential for AI-native preparation, planning, and operation phases in the network
network slicing in 6G networks.
slicing lifecycle.
Index Terms—6G, AI-native, network slicing, AI for slicing, Developing network slicing schemes faces many challenges
slicing for AI, ubiquitous intelligence. in 6G networks due to their unique features. First, managing
slices over space, air, and ground network segments in the
SAGIN requires judicious coordination of heterogeneous net-
I. I NTRODUCTION work segments. Moreover, the 6G networks need to support
Compared with existing wireless networking including the a variety of new services while satisfying their different and
fifth generation (5G), 6G is more than an improvement of key stringent QoS requirements, which further complicates slice
performance indicators (KPI) requirements, such as increased management. Hence, it is paramount to develop intelligent
data rates, enhanced network capacity, and low latency. The 6G slice management solutions in 6G networks. Second, fuelled
networks are envisioned to have the following unique features. by powerful computing capability and advanced AI techniques,
First, space networks, e.g., low earth orbit (LEO) satellites, air ubiquitous intelligence is fostering abundant AI services with
networks, e.g., unmanned aerial vehicles (UAVs), and ground new QoS requirements, such as data quality, inference accu-
networks, e.g., cellular base stations (BSs), are integrated into racy, and training latency. Hence, it is necessary to construct
a space-air-ground integrated network (SAGIN) to provide customized network slices to support the emerging AI services
global coverage and on-demand services [1]. Second, re- in 6G networks.
source virtualization using network slicing techniques and end In this article, we propose an AI-native network slicing
user virtualization using digital twin techniques can facilitate architecture for 6G networks to facilitate intelligent network
advanced network virtualization to provide flexible network management while supporting emerging AI services. AI-native
management [2], [3].1 Third, intelligence penetrates every means that, as a built-in component in the network slicing
corner of networks, ranging from end users, the network edge, architecture, AI exists not only in the software-defined net-
to the remote cloud, which results in ubiquitous intelligence. A working (SDN) controller for managing network slices, but
number of network nodes are endowed with built-in artificial also in network slices as services for end users. Hence, the
intelligence (AI) functionalities, thereby not only facilitating synergy of AI and network slicing in the proposed architecture
intelligent network management but also fostering AI services, is two-fold: On one hand, AI techniques can be applied to
manage network slices, namely AI for slicing. The network
W. Wu, C. Zhou, M. Li, H. Wu, X. Shen, and W. Zhuang are with the slicing lifecycle including preparation, planning, and operation
Department of Electrical and Computer Engineering, University of Waterloo, phases is introduced, along with specifying AI-based solutions
Waterloo, ON N2L 3G1, Canada (email:{w77wu, c89zhou, m475li, h272wu,
sshen, wzhuang}@uwaterloo.ca). Corresponding author: Mushu Li; for each phase. In addition, the detailed procedure of informa-
H. Zhou is with the School of Electronic Science and Engineering, Nanjing tion exchange among end users, access points, and the SDN
University, Nanjing 210023, China (email: haibozhou@nju.edu.cn); controller is presented; On the other hand, network slicing can
N. Zhang is with the Department of Electrical and Computer Engi-
neering, University of Windsor, Windsor, ON N9B 3P4, Canada (email: be applied to construct customized network slices for various
ning.zhang@uwindsor.ca). AI services, namely slicing for AI. Potential approaches such
1 End user virtualization is to virtualize end users in network operation
as AI instance construction and efficient resource management
and management by characterizing users’ behaviours and status (e.g., service
demands and QoS satisfaction), which can be achieved using digital twin for AI services are introduced.
concepts. The remainder of this article is organized as follows. In
2

Section II, expected features of 6G networks are discussed, and • Diversified services - Many services have stringent QoS
then the AI-native network slicing architecture is proposed. requirements in different dimensions. Mobile virtual re-
The basic ideas of AI for slicing and slicing for AI are pre- ality (VR) and hologram video streaming applications
sented in Section III and Section IV, respectively. A case study require a high data rate, e.g., the uplink data rate of
is presented in Section V. In Section VI, the research directions mobile VR is up to 5 Gbps. Other applications may
are identified, followed by the conclusion in Section VII. require ultra-high reliability, such as autonomous driving,
industrial control systems, and robot/UAV swarm, e.g.,
II. AI-NATIVE N ETWORK S LICING FOR 6G N ETWORKS the required reliability of autonomous driving is up to
99.999% [10];
A. Network Slicing • Ubiquitous intelligence - With caching capability, a large
Network slicing is an emerging technology to support amount of data can be stored in the network. In addition,
diversified applications in a cost-effective manner [4], [5]. with the development of AI techniques, edge computing,
The concept of network slicing can be traced back to the and device computing, intelligence is pushed from the
late 1980s [6]. Nowadays, network slicing is a key technology remote cloud to the network edge and end users. As such,
in 5G networks, supported by network function virtualization AI will be integrated into 6G networks for intelligent
(NFV) and SDN techniques. Specifically, NFV enables virtu- network management by directly learning from extensive
alized resources and network functions for flexible resource data in the network. Moreover, ubiquitous intelligence
management, while SDN facilitates centralized network man- will foster a number of AI services in which AI is
agement for network optimization. In 5G networks, network provided as services.
slicing has been defined in the 3rd generation partnership
project (3GPP) Release 15 [7]. Moreover, in coming 6G
networks, network slicing will continue evolving and play an C. AI-Native Network Slicing Architecture
increasingly important role. These features impose new challenges on developing net-
The basic idea of network slicing is to create multiple work slicing schemes for 6G networks. Firstly, the SAGIN
logically-isolated network slices on top of the common phys- not only increases the number of integrated network segments
ical infrastructure, which can achieve flexible and adaptive but also introduces extra dynamics on network resource avail-
network management. Its benefits are three-fold: 1) Multi- ability due to satellite mobility and UAV manoeuvrability.
tenancy - Multiple virtual networks can share the common Network slicing schemes should accommodate for the large-
physical infrastructure, thus reducing capital expenditures in scale SAGIN while taking dynamic resource availability into
the network deployment; 2) Service isolation - Multiple slices account. Moreover, supporting diversified services with strin-
are constructed for different services via judicious resource gent QoS requirements further complicates network slicing
management, such that service level agreements of different scheme design. Secondly, ubiquitous intelligence facilitates
slices can be effectively guaranteed; 3) Flexibility - Network many emerging AI services which will be prevalent in 6G
slicing can support flexible network management, as slices can networks. Different from conventional services, facilitating AI
be created, modified, or deleted on-demand. services requires multiple steps, including collecting high-
quality data samples, training satisfactory AI models, and
performing low-latency model inference, which should meet
B. Features of 6G Networks
diverse QoS requirements. How to satisfy such diverse QoS
From 5G to 6G, it is in general expected KPI requirements requirements for AI services remains a challenging issue.
to be increased by at least an order of magnitude. According To address the above challenges, an AI-native network
to a recent white paper [8], the KPI requirements of the slicing architecture for 6G networks is presented. As shown
6G networks include 1 Tbps peak data rate, 20-100 Gbps in Fig. 1, the architecture aims at integrating SAGIN and
user experienced data rate, 0.1 ms end-to-end latency, 10 ubiquitous intelligence and supporting diverse services with
million devices/km2 , and near 100% coverage. Such KPI re- stringent QoS requirements. Compared with network slicing
quirements demand several candidate technologies, such as for 5G networks, the proposed architecture has two new
THz communications and AI [6]. The 3GPP working group characteristics. Firstly, AI is integrated into SDN controllers
will discuss 6G candidate techniques by the end of 2026, and to realize intelligent network slicing, such that a number
the first 6G standard is expected to debut by 2030. of network slices with stringent QoS requirements can be
Distinguished from 5G networks, 6G networks have several managed efficiently and cost-effectively via AI techniques,
features: which is referred to as AI for slicing. Secondly, emerging
• SAGIN - While current ground networks provide good AI services are supported by network slicing. In addition to
coverage in highly populated areas, 6G needs to provide network slices for conventional services, new network slices
universal coverage, including in rural areas, remote lands, are constructed for AI services on top of the common physical
and sparsely populated areas. To achieve this goal, 6G infrastructure, which is referred to as slicing for AI.
will exploit the altitude dimension. Space, air, and ground Two types of SDN controllers are deployed in the proposed
network segments are integrated into the SAGIN [1], [9], architecture. One is the centralized SDN controller located at
which can provide global coverage, facilitate on-demand the cloud, which is to manage network slices. The other is
services, and support high-rate low-delay services; the local SDN controller located at access points, which is
 3

Centralized SDN
Controller

High Reliability

Massive Connectivity

Low Delay

AI Service High Accuracy

Space and Air Network

Core Network

Ground Network

User with AI Service AI Function Switch Local SDN Controller

Satellite UAV Access Point

Fig. 1. An illustration of the AI-native network slicing architecture for 6G networks.

to schedule resources to end users within each network slice. service delay, service priority, throughput, and reliability.
The SDN controller in the following refers to the centralized The 3GPP has standardized specific service/slice type
SDN controller unless otherwise stated. In the following, we values for classified services, such as enhanced mobile
will illustrate the basic ideas of AI for slicing and slicing for broadband, ultra-reliable low-latency communications,
AI in Section III and Section IV, respectively. and massive machine-type communications services [7];
• Network resource and function virtualization - Net-
III. AI FOR S LICING work resources, such as communication, computing, and
In this section, we introduce the network slicing lifecycle caching resources, are pooled into virtualized resource
with three phases and then investigate potential AI solutions blocks via advanced resource virtualization techniques.
for each phase. Next, the corresponding procedure of infor- Similarly, network functions, such as firewall, network
mation exchange in AI for slicing is discussed. name translation, and domain name system, are sepa-
rated from dedicated hardware network functions into
A. Network Slicing Lifecycle virtualized network functions (VNFs). Through virtual-
ization, the SDN controller can flexibly manage network
The network slicing lifecycle consists of three phases: resources and functions.
preparation, planning, and operation, as shown in Fig. 2. The
centralized SDN controller is in charge of the preparation and Once these tasks are completed, the SDN controller can
planning phases, while the operation phase is coordinated by construct network slices for each admitted slice request.
local SDN controllers. 2) Planning Phase: This phase aims at reserving network
1) Preparation Phase: This phase is to construct and resources to slices for service provisioning. The planning
configure network slices based on service requirements, data phase operates in a large timescale. Time is partitioned into
traffic, user information, and virtual network resource avail- multiple planning periods (windows) for each slice. The dura-
ability. To achieve the goal, the SDN controller conducts the tion of each planning window depends on service demand and
following tasks: network dynamics, whose value ranges from several minutes
• Service requirement extraction - This task is to classify to several hours. To achieve the goal, the following two steps
services by extracting their QoS requirements, such as are conducted in the planing phase:
4

Communication Computing Caching

Resource Resource Resource

Delay Throughput Reliability Accuracy ...

Virtualized Network Resource Pool
Service Requirement Extraction

User Mobility
Firewall
Pattern

Channel Condition Service Demand Network Name Domain Name

Stochastics Pattern Translation System
Centralized SDN
Stochastic Network Controller Virtualized Network
Information Function
Network Slicing Feedback

Resource VNF Resource

RAT Selection
Reservation Placement Orchestration

Resource Resource Block

Communication
Resource User 1
Slice 1 Planning
Computing Decision
User 2
Slice 2 Resource

User 3
Slice N Caching
Resource
Time
Planning Operation

Fig. 2. An illustrative example for the network slicing lifecycle which includes preparation, planning, and operation phases.

• Service and network information collection - Benefiting

from the global control functionality of the SDN con- Preparation
Historical
troller, extensive network information can be collected Service Demand
from underlying physical networks, such as service de- Slice
Admission
mands, stochastic channel conditions, and user mobility
Resource ...
patterns. The collected information is utilized for the Availability ...

following resource reservation decision making; Planning ...

...
Resource
• Resource reservation - At the beginning of each plan- Reservation
AI
ning window, the SDN controller adjusts the amount of Planning
Decision
reserved network resources for each slice based on the User Mobility,
Channel Condition
monitored slice performance. The reserved virtualized Physical
Network Operation
network resources of each slice are mapped to the phys- Resource
Orchestration
ical network. At the end of each planning window, some
system information is fed back to the SDN controller,
such as resource utilization, system performance, and Fig. 3. The considered AI-based network slicing solution in which AI plays
service level agreement satisfaction. Based on the feed- different roles in preparation, planning, and operation phases.
back information, the SDN controller can adjust resource
reservation decisions to accommodate dynamic network
environments while guaranteeing QoS requirements. (RAT), determining user association with specific radio access
points, deciding proper protocol and associated parameters,
3) Operation Phase: This phase is to schedule the service and orchestrating resources among end users.
of a slice using the reserved resources for subscribed end
users. The operation phase works in a much smaller timescale
(e.g., 100 ms) than that in the planning phase. Specifically, B. Roles of AI in Network Slicing
under the coordination of the centralized SDN controller, Although network slicing can facilitate service provisioning,
local SDN controllers allocate network resources to end users managing a number of network slices incurs significant net-
in each slice according to their real-time data traffic. The work management cost, especially in 6G networks. As shown
operation decisions include selecting radio access technology in Fig. 3, AI-based network slicing is a potential solution in
 5

which AI plays different roles in different network slicing 4) The SDN controller runs AI-based planning algorithms
phases. to make decisions based on the collected service-level
AI for preparation: In the preparation phase, AI needs to information;
perform two tasks. 1) Service demand prediction - Based 5) The determined planning decisions are sent back to all
on historical data, service demand can be predicted via AI access points;
techniques, such as recurrent neural networks. Prior studies 6) Access points enforce the received planning decisions,
show that the service demand and resource usage of a slice e.g., reserving network resources for corresponding
can be accurately predicted [11]. The prediction results can slices;
be utilized for decision making in the planing phase. 2) Slice 7) End users in service report their real-time information to
admission - The SDN controller admits slices to maximize their associated access points, such as real-time service
network resource utilization considering resource availability demands, channel conditions, and task data sizes;
and service demands. As the slice admission decision is binary, 8) Access points run the AI-based operation algorithm to
this problem is deemed as an integer optimization problem. allocate resources for end users based on real-time user-
In large-scale networks with complex resource availability level information;
distribution, conventional optimization solutions become com- 9) Service requests from end users are supported with the
plicated and intractable, while AI-based solutions are potential. allocated network resources. For example, computation
AI for planning: In the planning phase, AI can perform two tasks can be offloaded to access points using commu-
tasks. 1) VNF placement - The SDN controller deploys VNFs nication resource and then processed using computing
to support services in the network. The resources allocated for resources. For each operation slot within a planning
VNFs should be dynamically adjusted for time-varying service window, Steps 7-9 are repeated;
demands to guarantee service delay requirements. Deep learn- 10) Access points monitor slice performance in the network
ing methods can be applied to enhance resource utilization given the enforced planning decisions by measuring end
in dynamic network environments. 2) Resource reservation users’ satisfaction rates across all operation slots within
- The SDN controller reserves resources for different slices a planing window;
based on their service demands. Since data traffic loads are 11) Access points report network performance to the SDN
time-varying, the resource reservation should be adaptive to controller;
dynamic real-time demands, which can be addressed via rein- 12) The SDN controller makes the planing decision for next
forcement learning (RL) methods, such as deep deterministic planing window and adjusts the planning policy based
policy gradient (DDPG). on the feedback information.
AI for operation: Two exemplary operation tasks are as In the preceding procedure, Steps 1-12 are in the network
follows: 1) Resource orchestration - The reserved resources planning phase, and Steps 7-9 are in the operation phase.
of a slice are allocated to end users. The decisions are
determined based on real-time user mobility, service demands, IV. S LICING FOR AI
etc. To efficiently utilize resources, RL methods can be applied The slicing for AI is to utilize network slicing to support AI
for dynamic resource orchestration; 2) RAT selection - To services while satisfying QoS requirements. Potential solutions
maximize system utility, an optimal RAT is selected among include constructing and selecting AI instances and efficient
multiple candidate RATs for each end user. Due to user resource management in the AI service lifecycle.
mobility, user-perceived service performance of an RAT is
stochastic. Such problem can be addressed by multi-armed
A. AI Instance
bandit methods, e.g., contextual bandit.
There are diversified implementation options for supporting
AI services. An AI service can be implemented via different
C. Procedure of Information Exchange kinds of algorithms, training manners, and network resource
allocation. For example, objective detection services can be
The AI for slicing procedure involves the information implemented via ResNet32, Inception-v3, AlexNet, or VGG16
exchange among end users, access points, and the SDN algorithms. Hence, the primary issue of supporting an AI
controller. The procedure is illustrated in Fig. 4 with steps service is to determine an appropriate implementation option
as follows: in the network.
1) Access points collect user-level stochastic information, We introduce the concept of AI instance to address the
such as end users’ service demand patterns, mobility issue, as shown in Fig. 5. An AI instance of an AI service
patterns, and stochastic channel conditions; represents an implementation option for an AI service. The
2) Access points translate the user-level information into basic idea is to construct multiple candidate implementation
desired service-level information. For example, user options and then select an appropriate one based on network
density information can be obtained from processing environments. The procedure of the conceptual AI instance
user location information, and AI techniques can be used management framework consists of two steps. 1) AI instance
for such data abstraction, fusion, and analysis; construction - The network operator constructs multiple can-
3) The processed service-level information is delivered to didate AI instances for each AI service based on available
the SDN controller; virtualized network resources and AI service requirements.
6

End Users Access Points SDN Controller

1. User-Level Stochastic Information

2. AI-Based Information
Processing

3. Service-Level Information

4. AI-Based Planning
Decision Making

5. Planning Decision
Planning
6. Planning Decision Window
Enforcement

7. User-Level Real-Time Information

8. AI-Based Operation
Operation
Decision Making
Slot
9. Data

Repeat 7-9 During Each Operation Slot

10. Slice Performance

Measurement and Monitoring
11. Slice Performance

12. Adjust AI-Based Planning

Policy

Fig. 4. Procedure of information exchange in AI for slicing.

An AI instance may include (i) the AI algorithm which collaboratively to train a global model via federated learning.
specifies the implementation algorithm and the corresponding Next, well-trained AI models are deployed to execute specific
neural network architecture, (ii) the training manner of the computation tasks, which is referred to as model inference. The
AI algorithm, e.g., centralized or distributed training, and model inference can be performed in multiple manners. For
(iii) the amount of the required network resources. 2) AI example, device-edge collaborative inference approaches can
instance selection - In this step, the AI service provider selects allocate and process computation tasks at different network
an appropriate AI instance. For this purpose, the AI service nodes to achieve a low inference latency.
provider observes the status and collects service requirements The performance of an AI service depends on all the
of its current subscribed end users, and then selects an AI three stages in the AI service lifecycle. For example, model
instance among candidate AI instances provided by the net- inference accuracy depends on multiple factors, such as the
work operator. If an AI instance is selected, the AI service quality of the collected data, the number of training iterations,
will be executed using the AI algorithm and the corresponding and the approach of model inference. Meanwhile, all these
required amount of network resources by the AI instance. In three stages consume multi-dimensional network resources. As
summary, the idea of AI instance provides flexibility for AI a result, to optimize the performance of AI services, network
service management. resources should be jointly allocated for these three stages.
The reserved network resources in AI slices should be further
B. Resource Management in AI Service Lifecycle allocated to these three stages to satisfy their corresponding
Running AI services includes three stages: data collection, QoS requirements.
model training, and model inference, i.e., AI service lifecy-
cle [12], [13]. Specifically, data collection is to collect data
via communication links, and the collected data can be stored V. C ASE S TUDY
in network edge servers. Based on the collected data, an
AI model can be trained in the model training stage. The In this section, a case study is provided on AI-assisted
model training can be implemented in either a centralized or resource reservation, aiming at reducing long-term overall
a distributed manner. For example, multiple devices can work system cost.
 7

...
User

Extract
User Information

... AI Service
Provider
AI Algorithm Select
AI Instance
Training Manner

Computing Resource
AI AI AI AI AI AI
Instance Instance Instance Instance Instance Instance
Communication Resource ... ... ... ...

Caching Resource

Construct
AI Instance

Network
Operator
Virtualized Network Resource Pool

Fig. 5. The conceptual AI instance management framework for AI services.

A. Considered Scenario delay requirement. These weight parameters are set to ωr = 1,

ωs = 20, and ωd = 200, respectively. The planning window
We consider an air-ground integrated network for providing size is set to one hour.
autonomous driving services to vehicles traversing a highway We propose a DDPG-based solution to minimize the overall
segment. For the considered highway segment with a length system cost [15]. In this solution, both actor and critic net-
of 2 km, two BSs are uniformly deployed along the highway works are fully-connected neural networks with four layers,
with a separation distance of 1 km, and one UAV is deployed the numbers of neurons in two hidden layers are 128 and 64,
in the centre hovering at a height of 100 meters. When an respectively, and their learning rates are set to 2 × 10−4 and
autonomous vehicle is driving on the highway, extensive 2×10−3 , respectively. For performance comparison, we adopt
computation-intensive tasks are required to be processed. For an optimization-based solution, named myopic resource reser-
prompt task processing, all access points are equipped with vation, in which network resources are reserved to minimize
edge computing servers, and vehicles can associate to the the resource reservation cost at each planning window while
nearest access point and upload their computation tasks. We satisfying the delay requirement.
consider a delay-sensitive autonomous driving service, e.g.,
object detection, whose delay requirement is 100 ms for road
B. Simulation Results
safety [14]. The data size and computation intensity of a task
are set to 0.6 Mbit and 6×108 cycles, respectively. The service We evaluate the performance of the proposed DDPG-based
delay is characterized by the queuing theory since task arrivals solution based on real-world highway vehicle traffic flow trace
are assumed to follow a Poisson process with rate λ = 1 collected by Alberta Transportation.2 As shown in Fig. 6(a),
packet/sec. To guarantee the service delay requirement, a we first present the convergence performance of the proposed
network slice is constructed, in which spectrum and computing DDPG-based solution. A five-point moving average is applied
resources are reserved. to process raw simulation points to highlight the convergence
trend (i.e., red curve). It can be seen that the DDPG-based re-
Resource reservation decisions are made to minimize
source reservation solution has converged after 4,000 training
the overall system cost considering vehicle traffic dy-
episodes.
namics. The overall system cost is defined as C =
PT t t t Next, as shown in Fig. 6(b), the cumulative system cost
t=1 (ω r Cr + ωs Cs + ωd Cd ), which is a weighted summa-
within one day is presented. It can be observed that the DDPG-
tion of three cost components across all T planning windows.
based solution can reduce the cumulative overall system cost
1) Resource reservation cost Crt accounts for the amount of the
within one day by around 15% as compared to the myopic
reserved spectrum and computing resources at BSs at planning
solution. The reason is that the proposed DDPG-based solution
window t. The spectrum resource is allocated in a unit of
is able to minimize the long-term overall system cost, while the
subcarrier of 5 MHz, and the computing resource is allocated
myopic solution minimizes the short-term system cost, which
in a unit of virtual machine (VM) instance with a processing
incurs prohibitive slice reconfiguration cost due to frequent
rate of 10 × 109 cycles/sec; 2) Slice reconfiguration cost Cst
adjustment of network resource reservation in highly dynamic
accounts for the difference between two consecutive resource
vehicular networks. The simulation results show that the
reservation decisions [15]; 3) Delay requirement violation
penalty Cdt refers to the penalty once service delay exceeds the 2 Alberta Transportation: http://www.transportation.alberta.ca/mapping/.
8

3000
data is widely distributed in the network. Due to limited
DDPG Solution communication resources, the cost of collecting a large amount
of data cannot be neglected. In addition, the collected data
2500
is required to be processed to mine valuable information
Overall System Cost

for network management. For example, abundant historical

2000 behaviour data from individual users can be analyzed to
predict spatio-temporal service demand distributions. Hence,
1500 establishing a data management framework to collect and
analyze data is necessary.
1000
C. Prediction-Empowered Network Slicing
With the development of advanced AI technologies, the data
500
0 1000 2000 3000 4000 traffic in the network can be predicted. How to effectively
Training Episodes leverage the power of prediction for network slicing is an in-
(a) Convergence performance teresting topic. Since the prediction is imperfect, the prediction
error may degrade the performance of network slicing. How to
800 evaluate the impact of prediction errors on system performance
Myopic Solution and to develop corresponding solutions are important research
700 DDPG Solution
15% issues.
Cumulative System Cost

600

500 VII. C ONCLUSION

400
In this article, we have proposed the AI-native network
slicing architecture to facilitate intelligent network manage-
300 ment and support AI services in 6G networks. The architecture
200 aims at enabling the synergy of AI and network slicing. The
AI for slicing is to help reduce network management com-
100 plexity, while adapting to dynamic network environments by
0 exploiting the capability of AI in network slicing. The slicing
0 5 10 15 20 25 for AI is to construct customized network slices to better
Time (Hour)
accommodate various emerging AI services. To accelerate the
(b) Cumulative system cost within one day pace of AI-native network slicing architecture development,
Fig. 6. Performance evaluation of the proposed DDPG-based resource extensive research efforts are required, such as in the identified
reservation solution. research directions.

ACKNOWLEDGEMENTS
proposed AI-based resource reservation solution can achieve
a low system cost. This work was financially supported by Natural Sciences
and Engineering Research Council (NSERC) of Canada. The
authors would like to thank Jie Gao, Qihao Li, and Kaige Qu
VI. O PEN R ESEARCH I SSUES
for many valuable discussions and suggestions throughout the
In the following, we discuss some open research issues work.
pertaining to AI-native network slicing.
R EFERENCES
A. Joint Design of Network Planning and Operation [1] N. Zhang, S. Zhang, P. Yang, O. Alhussein, W. Zhuang, and X. Shen,
Planning and operation are performed and coupled in two “Software defined space-air-ground integrated vehicular networks: Chal-
lenges and solutions,” IEEE Commun. Mag., vol. 55, no. 7, pp. 101–109,
different timescales. Specifically, the planning phase is per- 2017.
formed at a large timescale (e.g., minute level) to reserve [2] X. Shen, J. Gao, W. Wu, M. Li, C. Zhou, and W. Zhuang, “Holistic
resources for different slices based on service demands, while network virtualization and pervasive network intelligence for 6G,”
submitted to IEEE Commun. Surveys Tuts., 2021.
the operation phase is performed at a small timescale (e.g., [3] R. Minerva, G. M. Lee, and N. Crespi, “Digital twin in the IoT context:
sub-second level) to allocate the reserved resources to on- A survey on technical features, scenarios, and architectural models,”
demand users within each slice. Achieving the optimal net- Proc. IEEE, vol. 108, no. 10, pp. 1785–1824, Oct. 2020.
[4] X. Shen, J. Gao, W. Wu, K. Lyu, M. Li, W. Zhuang, X. Li, and J. Rao,
work slicing performance requires a joint optimization design “AI-assisted network-slicing based next-generation wireless networks,”
of planning and operation. IEEE Open J. Veh. Technol., vol. 1, no. 1, pp. 45–66, 2020.
[5] W. Zhuang, Q. Ye, F. Lyu, N. Cheng, and J. Ren, “SDN/NFV-empowered
future IoV with enhanced communication, computing, and caching,”
B. Data Management Framework Proc. IEEE, vol. 108, no. 2, pp. 274–291, 2020.
[6] X. You et al., “Towards 6G wireless communication networks: Vision,
The cornerstone of AI-native network slicing is abundant enabling technologies, and new paradigm shifts,” Sci. China Inf. Sci.,
data that can be used for AI model training. In 6G networks, vol. 64, no. 1, pp. 1–74, 2021.
 9

[7] A. Kaloxylos, “A survey and an analysis of network slicing in 5G

networks,” IEEE Communications Standards Magazine, vol. 2, no. 1, Ning Zhang xxx
pp. 60–65, 2018.
[8] N. Rajatheva et al., “White paper on broadband connectivity Weihua Zhuang (M’93-SM’01-F’08) has been with the Department
in 6G,” arXiv:2004.14247, 2020, [Online]. Available: of Electrical and Computer Engineering, University of Waterloo,
https://arxiv.org/abs/2004.14247.
[9] C. Zhou, W. Wu, H. He, P. Yang, F. Lyu, N. Cheng, and X. Shen, “Deep
Waterloo, ON, Canada, since 1993, where she is currently a professor
reinforcement learning for delay-oriented IoT task scheduling in space- and a Tier I Canada Research Chair in Wireless Communication
air-ground integrated network,” IEEE Trans. Wireless Commun., vol. 20, Networks. Dr. Zhuang is a Fellow of the Royal Society of Canada,
no. 2, pp. 911–925, 2021. the Canadian Academy of Engineering, and the Engineering Institute
[10] C. Campolo, A. Molinaro, A. Iera, and F. Menichella, “5G network of Canada. She is also an elected member of the Board of Governors
slicing for vehicle-to-everything services,” IEEE Wireless Commun., and VP-Publications of the IEEE Vehicular Technology Society. She
vol. 24, no. 6, pp. 38–45, 2017. was a recipient of the 2017 Technical Recognition Award from the
[11] C. Gutterman, E. Grinshpun, S. Sharma, and G. Zussman, “RAN IEEE Communications Society Ad Hoc & Sensor Networks Technical
resource usage prediction for a 5G slice broker,” in Proc. ACM MobiHoc, Committee.
Catania, Italy, 2019.
[12] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT press,
2016. Xuemin (Sherman) Shen (M’97-SM’02-F’09) is currently a Uni-
[13] M. Li, J. Gao, C. Zhou, X. Shen, and W. Zhuang, “Slicing-based AI versity Professor with the Department of Electrical and Computer
service provisioning on network edge,” IEEE Veh. Technol. Mag., 2021, Engineering, University of Waterloo, Canada. His research focuses
to appear. on network resource management, wireless network security, social
[14] S.-C. Lin, Y. Zhang, C.-H. Hsu, M. Skach, M. E. Haque, L. Tang, networks, 5G and beyond, and vehicular ad hoc networks. He is a
and J. Mars, “The architectural implications of autonomous driving: Canadian Academy of Engineering Fellow, a Royal Society of Canada
Constraints and acceleration,” in Proc. ASPLOS, 2018, pp. 751–766. Fellow, and a Chinese Academy of Engineering Foreign Fellow. He
[15] W. Wu, N. Chen, C. Zhou, M. Li, X. Shen, W. Zhuang, and X. Li, received the R.A. Fessenden Award in 2019 from IEEE, Canada; the
“Dynamic RAN slicing for service-oriented vehicular networks via
constrained learning,” IEEE J. Sel. Areas Commun., vol. 39, no. 7, pp.
James Evans Avant Garde Award in 2018 from the IEEE Vehicular
2076–2089, 2021. Technology Society; and the Joseph LoCicero Award in 2015 and
Education Award in 2017 from the IEEE Communications Society.

B IOGRAPHIES

Wen Wu (S’13-M’20) earned the Ph.D. degree in Electrical and

Computer Engineering from University of Waterloo, Waterloo, ON,
Canada, in 2019. He received the B.E. degree in Information En-
gineering from South China University of Technology, Guangzhou,
China, and the M.E. degree in Electrical Engineering from University
of Science and Technology of China, Hefei, China, in 2012 and
2015, respectively. Starting from 2019, he works as a Post-doctoral
fellow with the Department of Electrical and Computer Engineering,
University of Waterloo. His research interests include millimeter-
wave networks and AI-empowered wireless networks.

Conghao Zhou (S’19) received the B.Sc. degree from Northeastern

University, Shenyang, China, in 2017, and the M.Sc. degree from the
University of Illinois at Chicago, Chicago, IL, USA, in 2018. He is
currently pursuing the Ph.D. degree with the Department of Electrical
and Computer Engineering, University of Waterloo, Waterloo, ON,
Canada. His research interests include space-air-ground integration
networks and machine learning in wireless networks.

Mushu Li (S’18) received the B.Eng. degree from the University

of Ontario Institute of Technology (UOIT), Canada, in 2015, and
the M.A.Sc. degree from Ryerson University, Canada, in 2017.
She is currently working toward the Ph.D. degree in electrical
engineering at University of Waterloo, Canada. She was the recipient
of Natural Science and Engineering Research Council of Canada
Graduate Scholarship (NSERC-CGS) in 2018, and Ontario Graduate
Scholarship (OGS) in 2015 and 2016, respectively. Her research
interests include the system optimization in VANETs and machine
learning in wireless networks.

Huaqing Wu (S’15) received the B.E. and M.E. degrees from Beijing
University of Posts and Telecommunications, Beijing, China, in 2014
and 2017, respectively. She is currently working toward the Ph.D.
degree at the Department of Electrical and Computer Engineering,
University of Waterloo, Waterloo, ON, Canada. Her current research
interests include vehicular networks with emphasis on edge caching,
resource allocation, artificial intelligence (AI)/machine-learning (ML)
for future networking, and space-air-ground integrated networks.

Haibo Zhou xxx

View publication stats