1

5G Edge Automation & Intelligence| November 2021

2 5G Edge Automation & Intelligence| November 2021

Contents

Executive Summary ...............................................................................3

1. 5G Edge Automation .........................................................................4
1.1 Background ................................................................................5
1.2 State of the Art and Industry Landscape .................................7
1.3 Envisioned Features and Key Technologies..............................10
1.4 Requirement Analysis ................................................................12
1.5 Architecture Direction ................................................................13
1.6 System Recommendations for ML-driven Automation .............14
2. 5G Edge Optimization, Intelligence, and Analytics ..........................16
2.1 Background ................................................................................16
2.2 State of the Art and Industry Landscape ..................................17
2.3 Envisioned Features and Key Technologies .............................19
2.4 Requirement Analysis ................................................................24
2.5 Architecture Direction ................................................................25
2.6 System Recommendations for ML-driven Optimization ...........26
3. Application of 5G Edge Automation and Edge Intelligence ............29
3.1 Autonomous Industrial Solutions...............................................29
3.2 Intelligent Transport Systems ....................................................30
3.3 Smart Energy and Smart Homes...............................................31
3.4 Connected Health ......................................................................31
3.5 Enabling Location Information ..................................................32
3.6 Cloud and Edge Gaming ............................................................32
3.7 Scalable Digital Twin...................................................................32
Conclusion .............................................................................................34
Acronyms............................................................................................35
References.........................................................................................37
Acknowledgments..................................................................................40
3 5G Edge Automation & Intelligence| November 2021

Executive Summary

5G is the very first mobile wireless technology that seeks to connect anything anywhere, making communications both
ubiquitous and as transparent as the air we breathe. A 5G ecosystem is emerging that is using pervasive, ever-present
communications to expand the reach and scope of today’s human-centric consumption patterns. One of 5G’s greatest
challenges ahead involves an increasingly complex, multi-dimensional, multi-variable world in terms of an ever-expanding
multitude of service requirements and disparate traffic profiles. Additionally, a newly-introduced heterogenous mix
of enabling technologies for both communications and compute, as well as their control, are adding to this emerging
challenge.
An important part of the 5G ecosystem is edge computing. It facilitates radical new use cases that extend from the
data center core to the network edges. Edge computing allows for compute and analytics to be moved closer to the
data instead of exchanging ever growing amounts of data among cloud servers. As the ecosystem evolves, 5G and edge
computing will further converge to enable edge network management, collect and capitalize on massive amounts of data
while maintaining integrity and even ownership, and build pervasive intelligence for enabling various latency-sensitive,
enterprise, and private services.
Two tasks are fundamental to address the complexity challenge and for the ecosystem to succeed: automation
and optimization. Automation is required to cope with new scenarios in network and edge planning, operation, and
management. Automation should ease the operation of network and compute infrastructure for rapidly growing vertical
industries, transportation, and enterprise use cases that are bringing along with them new infrastructure owners.
Optimization should also ease the extension of cloud computing and fast-growing Artificial Intelligence/Machine Learning
(AI/ML) applications to the edge, as well as introduce self-optimization to best serve applications. Optimization strategies
will ensure that every node in mobile networks can provide low latency, high reliability, and pervasive intelligence
capabilities.
The first three chapters of this technical whitepaper details where and how intelligence can support the cross-over of
communications in a 5G network and edge computing in a 5G edge network. Chapter 1 focuses on automation, while
Chapter 2 looks at optimization of compute and communications. Chapter 3 provides insights on how to apply them.
The first two chapters will follow the same story line, starting with the specific background and introduction of the current
state of the art and industry landscape. This is followed by a discussion on features and key technologies. Based on this
discussion, the sections develop foreseeable requirements, illustrate emerging directions for network architectures, and
conclude with system recommendations. Chapter 3 combines this understanding, with applications in potential use cases
for autonomous industrial solutions, smart transport and energy, connected health, and digital twins – and more.
Core to the discussion include the technical capabilities of 5G and edge computing, where the intelligence of 5G network
and edge computing can achieve interesting results. These include the automation of data collection, analysis and
communication, and computing automation for ease of management. They may also include the optimization of network
and computing to best support AI/ML applications at the edge, enabling pervasive intelligence at connected devices,
distributed learning, situational networks, and collaborative edge intelligence.
This paper is the joint effort of experts from many different technical disciplines spanning research, development,
operations, and applications. The result is a comprehensive discussion and guide that also shows how multi-dimensional
complexity challenges can be tackled, bringing together expertise from multiple backgrounds united by a common
goal: automating and optimizing 5G networks to capitalize on edge computing advances that serve future customer
requirements and applications.
The combination of 5G and edge computing will create new capabilities and new business opportunities for the whole
communication and computing industry. We hope this white paper will help the ecosystem achieve a consensus around
the future of 5G and the edge technology roadmap, working closer to realize the vision, and bring a better automated,
intelligent world together.
4 5G Edge Automation & Intelligence| November 2021

1. 5G Edge Automation

Cellular services have historically been delivered from a centralized location.

Traffic from many cell sites is transported back to a central location, where it is
hosted in a mobile packet core that provides connectivity to the wider Internet.
However, the drawbacks of this approach have become apparent in the face of
emerging applications that require low latency, high reliability, high bandwidth,
and are characterized by localized communication among peers that share
a common local domain. In many cases, rather than centralizing the entire
service delivery, computation is needed for things such as automation of edge
applications like video analytics, process control, self-driving cars, and more.
In this context, mobile network virtualization and cloudification of Radio Access
Networks (RAN) have gained momentum in recent years. By decoupling
hardware and software, network element protocol stacks can utilize the
computing resources at the base stations or data centers near the radios and
cell towers. This new paradigm creates new pools of computing resources at the
edge of the network that can be utilized by other applications for different use
cases, which has caused edge computing to become successful over the past
decade.
The edge is defined as a continuum of edge zones, including such examples as:
• Device edge – signal and data processing on the device
• Premise edge – processing that occurs on the premise (home, car,
enterprise, etc.)
• Access edge – processing at cell sites or access Points of Presence (POP)
• Metro edge – upstream aggregation centers like Internet service providers,
etc.
This combination of different sizes of cloud data centers at global, national,
local/regional, and potentially access locations are integrated into the network
and operated by a central orchestration and management system. The exact
specification of the infrastructure on the different sites may depend on the
use cases and applications onboarded. In addition, there can be several
infrastructure providers at the same site.
Automation at each of these edge zones has different forms and requirements.
For example, an Internet of Things (IoT) device can perform autonomous and
intelligent local computation based on its sensed environment. Automation can
also impose a power reduction of IoT devices to conserve as much energy as
possible. Another example involves self-driving cars, which can process data
at the premise edge, as well as upload training data to the metro edge via the
access edge for machine learning algorithms. Hence, various edge zones can be
located at different locations, can execute different functions and/or decisions
for different requirements, and may exchange information or data with each
other.
5 5G Edge Automation & Intelligence| November 2021

1.1 Background
The term “Network Automation” or in short “Automation” has been used in the
communication industry to describe a wide range of technologies that would
help automate network system processes and service delivery with reduced or
minimized human intervention. Human intervention is reduced by introducing
predetermined decision criteria and related actions and embodying those
predeterminations into the processes. Automation includes the use of various
control systems for operating the network and services. Within this context,
terms like “policy”, “control-loop”, and “autonomic decision-making” are used.
A policy governs the choices behavior (decisions) in a system. The functionality
of a system, its invariant part, is called its “mechanism”. Policies are the variant
part of a mechanism and are either static or non-static.
While static policies do not change over time or based on conditions, non-static
policies change at runtime. There are three different non-static policy types:
context -aware, adaptable, and adaptive. A context aware policy changes its
decisions using context from its inputs (events), its own context, or outside
information called external context. An adaptable policy changes its decisions
based on any external stimuli (including context), such as with a particular
configuration or parameters. An adaptive policy changes its decision-making
behavior based on internal stimuli, for instance history of decisions or learning
from previous decisions.
Adaptable and adaptive policies can also be based on Artificial Intelligence (AI)
techniques, which are then typically categorized as intelligent policies. A policy
has input and output interfaces, which are translated into interfaces towards
some triggering mechanism for input and some actioning mechanism for output,
both of which are outside the policy system. Consequently, one can connect any
policy with any triggering system (and concrete protocol and communication)
and any actioning system. Extending the interfaces for feedback is also possible.
This is important especially in closed feedback control loops and can help to
build self-stabilizing policy systems.
A closed-loop control system is an essential feature of the automatized work
process. In its simplest form, within a closed-loop control system, a controller
compares a measured output value of a system with a desired value. In case of
a mismatch between these values, the controller decides on the actions needed
to achieve the desired value by updating or changing its decision-making policy.
Hence, closed-loop control is the enabler of adaptive decision-making. This
process does not require any manual input or control and therefore, leads to
automation.
Closed-loop control ensures that a deviation from a desired value is mitigated
by updating a system’s policy, which is known as autonomic decision-making.
From here, policies adapt to evolving or varying changes in the system. This
adaptation can be realized by enhancing the capabilities of the policy itself or
by automatically re-authoring a policy. Three aspects are important to achieve
autonomic decision-making:
• A policy should use a decision-making approach rather than a decision
selection approach
• A policy should be able to use contextual information for its decision-
making approach
• A policy should be able to change its decision-making process at runtime.
Once these features are enabled in a dynamic system, automation can be
achieved, and human intervention can be minimized, which is the anticipated
goal of automation at the 5G network edge. As a result, automation in general
will continue to advance rapidly in areas where there are real and tangible
benefits (e.g., manufacturing, automotive, telecommunications, etc.). For
instance, integrating AI advances into automation will enrich the human
6 5G Edge Automation & Intelligence| November 2021

experience, such as allowing people to have enriched dialogue with personal

devices and obtaining advice and guidance. AI continues to make steady
progress in areas of speech recognition, decision making, and visual perception.
Automation is therefore needed to dynamically control (via control loops) and
optimize (based on autonomic decision-making with adaptive policies) the
heterogeneous networks of tomorrow. To support such automation, AI and
Machine Learning (ML) are key ingredients. Figure 1 below demonstrates how
autonomic decision-making can be impacted by external goals and context.

Figure 1 Autonomic decision-making driven by external goals and context.

There are some key advantages to automation in general. First, there will be
an overall improvement in efficiency. Software agents are good at executing
repetitive tasks to streamline production output, reduce or eliminate human
errors and deliver high quality of assurance. AI data, collected and learned at
the edge, focuses on the environment at that specific local edge and dynamic
policies would adapt accordingly at that specific edge location. Second,
shifting to automation and AI will mean technicians are working fewer hours in
dangerous conditions, which will decrease workplace injuries. Third, intelligent
automation will also provide interesting opportunities for workers to focus on
more complex and innovative tasks.
Additionally, automation at the edge can be used to target workloads that
are sensitive to stringent timing latencies. In this regard, context information
that is exchanged between devices and agents deployed at the edge can
help to optimize the workload-resulting in the reduction of completion time or
improvement of algorithm performance.
Automation at the edge can be further improved with 5G’s drastically increasing
capabilities in data transfer rate, capacity, device density, and reduced latency
and energy consumption. This powerful combination enables a wide range of
use cases that will mark a significant transformation in our lives. From smart
homes, smart cities, smart cars to Augmented Reality/Virtual Reality (AR/VR),
3D video, and e-health, the 5G network will become a ubiquitous and pervasive
layer that touches every aspect of our daily lives.
In the meantime, innovative industrial applications enabled by 5G are being
utilized to monitor, alert, diagnose and control activities across manufacturing,
energy, utilities, transportation, smart grid, security, and public safety. The use
of Software Defined Network (SDN), Network Function Virtualization (NFV), and
edge computing in 5G increases flexibility by dynamically scaling resources,
offering compute capability near the devices and exposing real time network
measurements for the introduction of automation at the edge as well as AI-
based decision-making.
Automation at the edge can facilitate new use cases that add value to the end
customer. Near-real time analytics at the edge can provide timely insights to
optimize end user performance by prioritizing radio resources automatically. If
the Service Level Agreement (SLA) deteriorates or starts drifting, new resource
management policies can be automatically created and applied to the system
7 5G Edge Automation & Intelligence| November 2021

without human intervention. Another example where automation at the edge

can be beneficial is to apply reinforcement learning for problems where the
modelling of the environment is challenging or even currently impossible for
instance, in network slice admission control or for problems where the optimal
decision-making policy is not known due to the lack of a master/genie and must
be intelligently approximated via trial and error.
Lastly, reinforcement learning algorithms need to evaluate the entire state
of network resources, which might lead to scaling problems with increasing
network size, and intelligently admit or reject these creation requests based on
knowledge developed through exploration and exploitation of different decision
options. Nevertheless, a superior level of “thinking” and “learning” will be
needed to complement network automation on a path to realize the zero-touch
network vision.

1.2 State of the Art and Industry Landscape

The RAN, Transport and Core ecosystem processes and technologies are deeply
rooted in many standards organizations. These bodies establish standards that
deal with automation and edge and include Zero-Touch Network and Service
Management (ZSM), NFV Management and Orchestration (MANO), and Multi-
Access Edge Computing (MEC). We provide a brief overview of the state of the
art in edge automation and related standards and technologies and highlight
3rd Generation Partnership Project (3GPP) standards, such as Self-Organizing
Networks (SON), Network Data Analytics Functions (NWDAF), as well as
industry efforts like O-RAN and Tele-Management Forum (TMF), the European
Telecommunications Standards Institute (ETSI) ZSM standard, and open-source
activities.

1.2.1 Edge Automation related 3GPP Standards

As 5G networks are intended to support various new services such as
IoT, cloud-based services, industrial control, autonomous driving, mission
critical communications, etc. with ultra-low latency and high data capacity
requirements, the 5G system architecture [1] supports edge computing
to enable such services by applications that are hosted closer to the user
equipment’s (UE) access point of attachment in order to reduce the end-to-end
latency and the load on the transport network. Additionally, edge computing
deployment scenarios and use cases have been defined to guarantee end-to-
end service requirements and discuss potential deployment solutions [2].
Furthermore, 5G system enhancement for edge computing capturing a reference
architecture, connectivity models, procedures for supporting edge computing [3]
as well as enhancements of edge computing management [4] have also been
introduced. Finally, from the application point of view, a technical specification
that provides an application layer architecture and related procedures for
enabling edge applications over 3GPP networks have been defined [5].
Moreover, studies on architecture enhancements for 5G systems to support
network data analytics services [6], enablers for network automation for 5G
systems [7], enhancement of Management Data Analytics (MDA) [8], [9],
enhancement for data collection for 5G New Radio (NR) and Evolved-Universal
Terrestrial Radio Access New Radio Dual Connectivity (EN-DC) [10], [11], [12].

1.2.1.1 Self-Organizing Networks

The concept of SON plays a major role and is an integral part of legacy mobile
radio access networks. SON is an automation technology to enable simpler
and faster planning, configuration, management, optimization, and healing of
the mobile network. SON is commonly divided into three architectural types:
centralized SON (C-SON), distributed SON (D-SON), and hybrid SON (see Figure
2). C-SON functions are typically concentrated closer to higher-order network
elements where network management systems are located to potentially allow
8 5G Edge Automation & Intelligence| November 2021

Figure 2 C-SON and D-SON control loops in O-RAN.

the autonomous control of a larger made based on network analytics, plane optimization and the interaction
number of network elements and their which allows the Policy Control between NWDAF and AI model and
coordination. In D-SON, functions are Function (PCF) or any other 5G core training services. Within a network,
distributed among various network Network Function that has subscribed NWDAF can be implemented in a
elements, including the edge, to to NWDAF output to perform decisions, centralized manner, distributed
potentially autonomously control such as to update and/or adapt a manner, or a hybrid of the two.
various SON features, such as load policy by considering the analytics When NWDAF is implemented in a
balancing or antenna tilt configuration information provided by the NWDAF. distributed or in a hybrid manner, it
in a closed loop environment. A hybrid is possible that distributed instances
The PCF request may be triggered
SON is a mix of C-SON and D-SON. of NWDAF are placed at the edge to
based on a request from other help with edge automation use cases.
With the emergence of enhanced AI/ network functions, modification In this manner edge data would be
ML techniques, enhancements of SON requests, or any changes in the stored, processed and analyzed
with cognitive features are gaining network. The following analytics locally. This would also help reduce
momentum with the introduction of are relevant for policy decisions: latency and the overhead of carrying
more agility in the network achieved “Load level information”, “Service data across the network.
by software/hardware decoupling, Experience”, “Network Performance”,
virtualization, increased compute, and “Abnormal behavior”, “User Equipment 1.2.2 Open Radio Access
decoupling of the protocol stack. While (UE) Mobility”, “UE Communication”,
Networks
research in this context is ongoing, “User Data Congestion”, “Data
novel concepts and solutions are Dispersion”, and “WLAN performance”. Open Radio Access Networks are
introduced and explored to handle As illustrated in Figure 3, such input often abbreviated as “Open RAN”,
the complexity of 5G networks with data and analytics are collected by “OpenRAN”, as well as “O-RAN”.
zero-touch network optimization and the NWDAF to make policy decisions. For the purposes of this white
real-time problem solving. The output of the NWDAF serves the paper, “Open RAN” (note the
network functions and the OAM to space between the words) refers to
1.2.1.2 NWDAF decide how to use the data analytics open and interoperable interfaces
On the 5G core network side, network provided by the NWDAF to improve the within and between various
automation and data analytics have network performance, which reflects a subcomponents of the RAN. Hence,
been enabled with the introduction closed-loop control system framework. it refers to the movement in wireless
of the NWDAF in 3GPP Release-15. telecommunications to disaggregate
NWDAF is expected to have a hardware and software and to create
These operations have also been distributed architecture providing
enhanced in subsequent releases open interfaces between them.
analytics at the edge. Currently,
[13]. NWDAF was introduced studies on network automation OpenRAN, on the other hand, refers
to provide analytics to 5G core enhancements are ongoing and focus to one of the two groups within the
network functions and to Operations on topics such as how to enable Telecom Infra Project (TIP), i.e., the
Administrations and Management real-time or near-real time NWDAF, OpenRAN project group, which is an
(OAM). Network policy decisions are how to enable NWDAF-assisted user initiative to define and build previous
9 5G Edge Automation & Intelligence| November 2021

Figure 3 General framework of 5G (core) network automation.

generations RAN solutions based These parts and their respective group that enables automation of the
on general-purpose, vendor-neutral interfaces are identified and clarified end-to-end network management with
hardware and software-defined in Figure 4. minimal to zero human intervention.
technology, or the OpenRAN 5G NR The scope of ZSM includes the RAN,
• Orchestrator and RIC component
project group which focuses on 5G NR. transport, core, NFV, SDN, legacy, and
– A1 interface. everything in between, that makes up
O-RAN (or alternatively ‘ORAN’ in • RIC and Centralized Unit/ a communication service. The ZSM
some cases), is an acronym for the Distributed Unit (CU/DU) – E2 framework facilitates collaborative
O-RAN Alliance, which publishes new Interface. management interactions between all
RAN specifications, releases open
• CU-CP (Control Plane) and CU-UP elements and all layers of the network
software for the RAN, and supports
(User Plane) – E1 Interface. enabled by closed-loop automation, AI,
its members integration and testing
• CU-DU – F1 interface. adaptive ML, and cognitive technology.
of their implementations. Throughout The architecture is service-based,
this white paper, the abbreviations will • DU-RU (Radio Unit) – Open modular, flexible, scalable, and is
be used based on these definitions. FrontHaul. defined in [17].
Open RAN disaggregates the RAN. • Orchestrator and Cloud Platform
(O-Cloud) – O2 Interface. The ZSM framework defines standards
In O-RAN this is done by using interfaces that enable interactions
open interfaces and incorporating 1.2.3 Tele-Management between management domains,
the concept of RAN Intelligent Forum (TMF) coordination between different closed
Controllers (RICs) that can host loops, and interactions between AI
smart applications (i.e., rApps and TMF has several projects focused on
components and closed loops, and
xApps) and perform radio resource edge automation. For example, the
hence, provides the glue allowing
management functions at a per UE Catalyst project [15] is constructing
interactions of various components,
level. These controllers extend new a standardized Edge Compute-as-
enabling the autonomous
management and control interfaces a-Service (ECaaS) for realizing zero
management of the end-to-end
to the RAN ecosystem (namely O1, touch edge solutions. Another project,
network.
O2, A1 and E2). Edge cloud servers AI Operations (AIOps) [16] is targeting
will typically host the Near-Real Time how AI can drive closed-loop service The complexity of the resources at
RIC while centralized data centers will assurance in communications service the 5G edge can be abstracted by
typically host the Non-Real Time RIC. provider’s network services. a management domain in the ZSM
Automation at the edge will involve architecture. The interactions between
both types of controllers, i.e., RICs, 1.2.4 ETSI Zero Touch Network management domains and the end-
and orchestration engines. O-RAN is and Service Management to-end service domain are defined by
defining or clarifying the usage of the (ZSM) ZSM. Management domains in the
interfaces between the different parts ZSM architecture allow the separation
ETSI aims to deliver a framework
of the RAN. of management concerns [18], and
within the ZSM industry specification
10 5G Edge Automation & Intelligence| November 2021

Figure 4 O-RAN architecture [14].

can consist of components in the users. The 5G Super Blueprint covers of the network to minimize the latency
3GPP RAN, 3GPP Core, Transport RAN, edge, and core for enterprises and maximize the bandwidth of such
domain, NFV components, O-RAN, and verticals. It spans a broad variety processing. Standards for collecting
Open Network Automation Platform of use cases including 5G, AI, edge operational and control data from the
(ONAP), etc. Infrastructure-as-a-Service/Platform- network (such as O-RAN Alliance’s O1
as-a-Service (IaaS/PaaS), IoT. The and E2 interfaces [14]) are a must-
1.2.5 Open Source goal is to offer flexibility to scale edge have, but platforms must also provide
There are several open-source related cloud services quickly, to maximize the the ability to implement customized
projects relating edge clouds and applications or subscribers supported normalization and machine-learning
automation. The Linux Foundation on each server, and to help ensure procedures to meet highly variable
(LF) is driving several initiatives the reliability of systems that must be and rapidly changing business needs.
to bring together an ecosystem of functioning at all times.
An important concern is how to meet
open-source items. These items the data collection and processing
range from RAN orchestrators, policy 1.3 Envisioned Features and requirements in a manner that retains
guided RAN controllers to mobile Key Technologies the ability of network operators and
packet core. LF aims to establish an other stakeholders to mix-and-match
open, interoperable framework for This section describes envisioned
features and key technologies required solutions from different vendors, as
edge computing powered by AI/ML. well as the open-source community,
Fostering cross-industry collaboration for the implementation of edge
automation and its enhancement as so that they retain ownership and
across IoT, Telecom, Enterprise and control of the data instead of being
Cloud ecosystems is key to deliver well as which can benefit from edge
automation. locked into proprietary database
value to end users. technologies. An open, high
At the 5G edge, an open-source 1.3.1 Distributed Data performance data streaming solution
software for carrier-scale edge such as Kafka [19], Pulsar [20], and
Collection, Normalization, and
computing applications that run in Rabbit-MQ [21] will have a role to play
Real-Time Processing in any such architecture, along with
virtual machines and containers to
support reliability and performance Big data management is a challenging data lake technologies that provide
requirements has been introduced. area of research. The problem easy on-ramps and off-ramps for
Within the LF community, LF Edge, becomes even more of a challenge collected data and co-located compute
has several projects targeted for the when the data must be collected and facilities such as Spark [22] that will
edge cloud. LF Networking is leading processed to produce control signals minimize the need to transport data
a community-driven integration and that are sent back to the network in from place to place when executing
proof of concept involving multiple near real-time. A mix of data streaming normalization activities or building
open-source initiatives to show end-to- technologies, in-memory and on-disk insights.
end use cases demonstrating various storage, and compute facilities will
implementation architectures for end need to be located close to the edge
11 5G Edge Automation & Intelligence| November 2021

Figure 5 Intelligent Automation solution architecture.

1.3.2 Context Discovery and common infrastructure comprised incorporating analytic and AI capability
Situational Awareness of physical and virtual resources. into edge automation process,
As the definition implies, a network critical insight can be extracted from
The ability to turn data from multiple slice supports end-to-end network network measurements and be used
sources into knowledge and exposing connectivity for end-users, humans, to generate optimal, dynamic slicing
it as actionable insights is one of and machines. Therefore, most of configurations allowing rapid service
the key elements of understanding the actions performed on the traffic deployment and providing a framework
behavior while applying it to build in a 5G network will take place at with business agility and flexibility.
situational awareness. This requires the edge, where content is created
using computer vision, enhanced With intelligent edge automation,
and consumed. Each slice is virtually
location services, network APIs operators can implement zero-
isolated from another and is designed
exposing congestion data, and touch slice lifecycle management,
according to the specific needs of the
connection availability. which includes automating the
application or end-user.
design, creation, modification, and
In addition to that, Application With network slicing being critical to monitoring of end-to-end network
Program Interfaces (APIs) pairing the successful delivery of 5G services, slices as well as the provisioning of
application/device needs with mobile and wholesale operators alike underlying resources to a slice, as
matching available edge locations should be able to plan, design, and and when required. The intelligent
are necessary to offload and activate thousands of customized edge automation solution should also
connect clients to best fitting edge network slices for their customers very support the scaling and orchestration
cloud nodes. This should be based quickly. They also should be able to of network resources for 5G Core,
on geography, compute, storage, modify and scale a slice up or down xHaul (combination of backhaul,
memory, and power requirements to address changing performance midhaul, and fronthaul), and RAN,
as well as other contextual insights. demands for optimized end-user along with the creation and operation
This bi-directional exchange of experiences. of network slices. Figure 5 illustrates
information should allow edge-enabled such an intelligently automated
applications to be context-aware yet 1.3.3.1 Intelligent Edge solution architecture.
location independent. Automation for 5G Slicing
Intelligent automation software is
Security policies can also be applied Intelligent, analytics-driven automation the key to the proper placement of
based on network and location is more than just automating manual Cloud-native Network Functions/
insights. For example, an autonomous processes. It is the ability to take Virtual Network Functions (CNFs/
car would turn on Virtual Private input from several sources, such as VNFs) within a mobile network and
Network (VPN) for a public Wi- the network itself, analyze that input enabling Mobile Network Operators
Fi connection in a shopping mall to generate actionable insights, and (MNOs) to maximize the utilization of
before sending telemetry data on an then execute upon them via intelligent network resources by re-allocating
identified fault versus when connected actions. This type of automation is unused resources to other slices. With
to a secure 5G slice on the road. required in the complex end-to-end its advanced analytics, AI, ML, and
setting of 5G architecture. Intelligent automated orchestration features,
1.3.3 Network Slicing and automation is required at the edge an intelligent software solution
Dynamics of the network where most of the enables the creation of a self-driving,
A network slice is a logical end- services will be provisioned, including self-healing, and self-optimizing 5G
to-end network defined over a network slices. Furthermore, by network with zero-touch capabilities.
12 5G Edge Automation & Intelligence| November 2021

1.3.3.2 Zero-Touch Network Slice time to market for service offerings, Devices (UEs and others), sessions
Life Cycle Management increases customer satisfaction, and (network and application) and users
shortens the time to revenue for the now consume resources such as
The massive number of network operator. compute, storage, sensing, services,
slices that will be required and the and applications close to their current
speed at which the services will have 1.3.3.4 An Open and Standards- location in the network (or at least
to be managed, make it impossible Based Solution definitely closer than before). This
for mobile and wholesale operators requires a very different approach
to accomplish this process manually. A 5G automation solution should
to automate resource management,
Zero-touch automation capabilities support a wide range of industry
while also maintaining “classic” (non-
are a necessity to efficiently manage standards initiatives, open-source
edge) usage scenarios:
the lifecycle of network slices. Figure 6 projects, and Open APIs from the
demonstrates what the network slice TMF, Mobile Ecosystem Forum Edge resource management: This is
lifecycle management would look like (MEF), Open Networking Foundation usually the first target for automation
in a 3GPP standardized network. (ONF) and others. Aligned with this and should cover all relevant zones
approach, a zero-touch slice lifecycle (device, premise, access, metro)
An intelligent edge automation management solution should and must guarantee that availability
software solution can automate the support the Communication Service and quality access to all resources
entire Lifecycle Management (LCM) of Management Function (CSMF), the for any given number of consumers
network slices and includes support Network Slice Management Function (devices, sessions, and users) active
for the Global System for Mobile (NSMF), and the Network Slice Subnet in those zones. In static scenarios,
Communications Alliance (GSMA) Slice Management Function (NSSMF) where neither the consumers nor the
Template for the initial design phase. standard, as detailed in the 3GPP resources or the network conditions
Using this solution, operators can specifications, also referenced in the change, “classic” automation
plan, design, and create new network ETSI NFV MANO framework. techniques should be sufficient. Any
slices, monitor and modify a slice dynamicity – for example devices
to meet Quality of Service (QoS) or In order to adhere to emerging 5G
moving in and out of edge zones,
customer requirements, deactivate it standards, increase flexibility, and
sessions migrate between devices,
when no longer required, and release support a multi-vendor network,
compute and storage being changed,
associated resources back into a an intelligent analytics-driven edge
applications being updated or
federated inventory system. automation solution should be
moved – can result in rather complex
designed and developed as a cloud-
scenarios.
1.3.3.3 Rapid Service Deployment native application built upon a
containerized, microservices-based Edge infrastructure: compute and
As 5G adoption increases across new
architecture. storage, or clouds, are heterogeneous
industry verticals, explosive growth in
the number of services is expected. hardware and software environments.
Users will demand faster deployment 1.4 Requirement Analysis Automation will need to be able
of new services that meet their to deploy, monitor, and repair
Automation itself and the techniques
specific end-to-end QoS requirements. applications and sessions using
to automate processes are not
An optimal 5G automation solution them on a wide range of clouds
different in the edge compared
featuring network-wide correlated simultaneously. Edge automation
to other places in the network.
analytics, automated orchestration, should be independent of the
Edge automation will require event
and zero-touch capabilities can help underlying cloud providing multi-cloud
processing, analytics, closed loop
operators reduce the time to plan, mechanisms.
control, and policies the same way
design, and deploy new services as Transport or Core. However, the Location-aware vs. location-
across the multi-layer, multi-vendor, circumstances at the edge will require dependent: One goal of edge
multi-domain network from weeks or a specific way of applying these computing is to provide resources
months to a few minutes. The rapid techniques, with some degree of at specific locations close to the
deployment capability reduces the uniqueness. consumers. However, said consumers

Figure 6 3GPP network slice lifecycle management [23].

13 5G Edge Automation & Intelligence| November 2021

can and will move and are probably time requirements and the edge e.g., 5G NR, private 5G, Open RAN
not interested in accessing these will have to cope with very fine- RIC etc. Such a framework should
specific locations but resources with grained requests. Edge automation integrate a wider architecture aiming
defined QoS. Edge automation should should create, maintain, and remove for end-to-end observability, control,
be location-aware where location resources based on many specific and optimization of Open RAN and the
defines the automation target yet not time requirements rather than a 5G edge ultimately extending towards
be location-dependent. few generic service classes or QoS transport and core. Note that Open
parameters. RAN is one of the many potential
Mobility management and application- implementation baselines for an edge
dependency: a core property of a
1.5 Architecture Direction automation framework.
mobile network involves facilitating
consumer and, to a certain degree, The telecommunications industry An edge automation and control
resource mobility. A common scenario is amid several transformational framework may include:
is a mobile UE connected (via “long” shifts, including the adoption of 5G
• An Open RAN RIC (intelligent
tunnels or slices) to relatively static technology as well as momentum in
RAN control),
and aggregated endpoints (e.g., the marketplace from closed, single
packet gateways in the Core at a few vendor RAN to ORAN—standardized, • An Intelligent Access Controller
central locations). At the edge, the software-defined interfaces that are (multi-mode access support,
connections are “short”, endpoints are open and interoperable. While these access control),
no longer aggregated centrally, and are first steps towards introducing • An Intelligent Core Controller
the nature of end points changes from automation and control into the (control at the core and transport
some sort of gateway to applications. 5G network, further (architectural) network e.g., SD-WAN),
Edge automation must cope with enhancements must be introduced to • An Intelligent Edge Controller
application dependency while achieve automation at the 5G edge. (main coordination and
supporting consumer and resource To this end, an indicative end-to-end optimization function),
mobility, for example, a hand-over of architecture is shown in Figure 7 that
• Edge Intelligence (for recursive
an application between different edge encompasses an edge automation
monitoring, intelligent decision-
zones to follow a connected UE. and control framework, service
making, automation, and
management and orchestration layer,
Fine-grained time-sensitivity: one optimization)
intelligent edge applications layer and
reason to use edge computing is to • A Time-sensitive communications
a distributed edge infrastructure layer.
support time-sensitive services with data plane that can be
QoS impossible to achieve otherwise. An Edge Automation and Control dynamically programmed as
There are many time-sensitive framework is envisioned to include driven by application needs
properties, low latency being one of various components and to support
them. Consumers will define their standardized networks and entities,

Figure 7 Edge automation and control framework.

14 5G Edge Automation & Intelligence| November 2021

Intelligent Edge Controllers with Time- a new and exciting marketplace of functionalities should be amenable
Sensitive Applications applications. Much like the plethora to automation and adapt to evolving
of apps that are available in our situations. To achieve reliable systems
The introduction of edge controllers cell phones today, the Open RAN and responsible resource consumption
for the end-to-end control and marketplace will provide a rich and scalability of those systems, we
automation to support various ecosystem for innovators to develop need ML solutions to both automate
emerging application including time- various RAN optimization applications. the design and provide fundamental
sensitive applications, are expected to Diverse types of intelligent edge support for growing ML applications.
provide the required degree of control, applications (rApps and xApps) can
management, and orchestration for 1.6.1 ML for Systems
be developed to optimize both at a
the 5G edge automation. However, macro cell level or surgically at a per Machine Learning for systems explores
such a framework will have its own UE level. Non-RT RIC applications how we can leverage machine learning
challenges. Two main challenges in (rApps) operate at a time granularity tools and advances to improve
this context are security and a new of 10s of seconds while near-real-time systems. Several approaches can
marketplace. RIC applications (xApps) operate at a help to automate and manage future
Security: Through enhanced visibility much faster timescale in the order of networks, including advanced 5G
of key interfaces over open interfaces, 10s of milliseconds. networks.
AI/ML powered tools provide Implementing Open RAN technologies
automated security analytics. As 1.6.1.1 Smart Edge Configuration
could spur innovation and potentially
Open RAN evolves to incorporate new Every edge device has its own
provide additional benefits of
capabilities such as real-time and capabilities and resource constraints.
increased flexibility, agility, and
non-real-time RIC that use pre-trained For example, a cellphone is battery
resilience in the RAN. Decoupling the
AI models, new threats emerge related powered (unless connected to a power
hardware and software of the RAN not
to algorithmic conflicts, adversarial source), whereas a smart refrigerator
only creates possible opportunities
attacks, and data exfiltration. Open is connected to a constant power
for new businesses, both small and
RAN is evolving towards providing real- supply but may have weaker hardware
large, to enter the market, but it also
time security at the edge. Traditional configuration and capabilities due
could decrease the probability of the
implementations might take additional to profitability considerations. Many
vendor lock-in that can occur. The
time and manual diagnosis to curb such devices may participate in
modular nature also encourages
such threats possibly leaving the joint data collection and processing
the development of “best-of-breed”
network exposed. tasks and even communicate among
solutions due to increased vendor
O-RAN, with its virtualization, competition. Finally, a disaggregated themselves. Each such device should
disaggregation, automation, and and open ecosystem could provide have automation capabilities to
intelligence, is expected to be a resilience and agility benefits. coordinate an expectations negotiation
complementary part of 5G’s broader for different tasks and assess its
progression to greater security: 1.6 System limitations and ability to execute a
given application or task.
• Open interfaces ensure Recommendations for ML-
interoperability of protocols and driven Automation For example, a server or an application
security features coordinator may produce sample
The capabilities introduced by application benchmarks for each
• Disaggregation establishes 5G make it possible to take ML device to execute. Then, according to
diversity of supply chain applications to the next level, providing collected benchmark results, an ML
• Cloud-native applications provide a better user experience and open model can be used to determine the
isolation new opportunities, for instance in suitable hardware configuration and
An Open RAN architecture specified healthcare, security, and finance functionality for each participating
by the O-RAN Alliance, builds upon applications. However, making those device. This may lead to an automatic
3GPP RAN specifications with advances means that the underlying configuration of devices, better use
additional interfaces and functions. system must support unprecedented of their resources, and better system
However, these additional interfaces amounts of data being constantly utilization.
and functions introduce additional collected and processed in a timely
security risks due to the expanded manner as well as large amounts 1.6.1.2 Smart Edge Monitoring
attack surface. Internal and external of computation must be executed
There are several challenges involving
attacks can exploit vulnerabilities in promptly on the edge devices or on
edge device monitoring, statistics
the network architecture and cloud edge zones (i.e., premise, access, or
collection, and different fault and
infrastructure, while 5G use cases metro).
anomaly detection needs.
have decreased risk tolerance. To achieve these goals, instead of
Statistics collection: When considering
Marketplace: The introduction of adhering to the traditional worst-
each device separately, storage
intelligent edge applications opens case planning, even the most basic
and communication are of major
15 5G Edge Automation & Intelligence| November 2021

concern. For instance, a device can devices’ ability to predict upcoming ML applications can off-load many
have automation capabilities to malfunctioning relies on lessons tasks and enable better resource
adjust the resolution of its samples learned from the past. Automating usage and scalability. For example,
(e.g., images), or to recognize what such failure predictions is a big step a system may expose interfaces
information is more critical and toward system reliability on a larger that will allow for efficient in-network
time-sensitive and therefore should scale with life-supporting and critical aggregation, privacy (e.g., differential
be immediately dispatched (e.g., applications. privacy and secure aggregation),
healthcare application). and even encoding, decoding, and
Predict Network Overload: Network down-sampling capabilities when an
Anomaly Detection: Automated overload is a known problem. For application can indicate that lossy
anomaly detection includes a example, an audience of tens and information suffices.
wide range of applications such as even hundreds of thousands of people
finance, surveillance, health care, may lose connectivity at a sports 1.6.2.2 Support for Distributed
intrusion detection, fault detection in event. Predicting such overloads at
Data Collection
safety-critical systems, and medical specific times and locations, either by
diagnosis. For example, anomalies the network infrastructure or the edge Collecting data from billions of devices
in network traffic could mean that a devices, can be used to issue notices and for different applications and use-
hacked device is sending out sensitive alarming users from the possibility of cases is a major challenge. Having a
data to an unauthorized destination; such an event. It may be necessary for human expert to fine-tune this process
anomalies in a credit card transaction some users and application to take for each use case is impractical.
could indicate credit card or identity this possibility into account and take A desirable design goal is to have
theft; and anomaly readings from measures accordingly. a system API that supports data
various sensors could signify a faulty collection in a way that allows users
behavior in hardware or a software 1.6.2 Systems for ML and applications to specify which data
component. Running ML applications over edge is relevant and collect only relevant
devices is highly challenging in terms data. For example, an application can
Edge devices are usually resource- indicate through an API metrics by
constrained in terms of compute, of resource consumption. With the
increase of ML applications’ demands which the system can perform in-
communication, and memory, while network data filtering and aggregation
anomaly detection applications and usages, available amounts of
data, and growth in the number of during the collection process.
usually require detection of anomalies
as fast as possible. Therefore, it is users, these challenges will only Another step in automated monitoring
of increasing interest to develop, increase. The ability to meet these and statistics collection is the
support and deploy resource-efficient challenges largely relies on how we coordination of such activities over
anomaly detection ML models on such operate and perform communication a set of devices. For example,
edge devices. 5G can offer improved and computation for ML applications. coordinated monitoring and statistics
cross-device connectivity which can Systems for ML explores how the collection of several devices may
be leveraged to improve anomaly system design itself can improve ML result in better information and
detection performance, for instance performance and resource usage to resource usage. Likewise, eliminating
by employing efficient distributed data performance tradeoffs by exploiting redundancy in data collection process
collection. domain knowledge. Automation of executed by many devices can improve
systems to offer the appropriate tools both bandwidth and computational
Predict Infrastructure Failure: resources.
Previously, failed infrastructure meant and configurations to ML applications
that edge devices might have lost is an essential step towards better ML
connectivity and application accesses practices.
for the failure or takeover duration,
1.6.2.1 Support for ML
introducing inconvenience and worse.
However, considering the extent of Applications
future uses and applications, such ML applications often require a lot
an infrastructure failure can result of communication, computation,
in disastrous scenarios, leading to and data collection. For example,
a huge financial damage (e.g., an in federated learning, participating
automatic order to sell or buy stocks) devices perform potentially
and even a loss of life (e.g., medical computationally-extensive operations
and life-supporting applications). on their local data, exchange
Keeping available backups ready to parameter updates, and must be
take over may prevent such scenarios. concerned with the privacy of their
A complimentary and scalable solution data.
is the ability to predict infrastructure
failure. A device or neighboring Exposing suitable system APIs to such
16 5G Edge Automation & Intelligence| November 2021

2. 5G Edge Optimization,
Intelligence, and Analytics

The world of computing and communication has been going through a

fundamental paradigm shift in recent years. Moore’s Law has increased
both computing and communication capacity of network nodes and devices
tremendously. The end user/edge devices are becoming more and more
compute-intensive to run sophisticated optimization approaches and/or AI/
ML workloads to drive actionable insights at the edge. At the same time many
new time-sensitive, mission critical services require distributed, collaborative
processing at near edge users as opposed to centralized/cloud-based
processing.
The combination of 5G connectivity and AI/ML computing capability at the edge
enables more intelligent applications for the network edge nodes and devices
as well. Devices with limited processing power can leverage 5G network edge
node resources or other more capable devices nearby to gain intelligence.
Network edge nodes can also leverage devices to collect real-time data from
devices’ sensors and create a joint perception of the environment. They can also
collaborate in sharing sensor data, AI/ML processing, and coordinate actions
thus enable collaborative intelligence.
This shift creates new business opportunities for 5G edge computing and
communications technologies as the ability to run AI/ML algorithms on the
edge infrastructure that are connected to and can serve edge users. Such edge
infrastructures are lucrative for circumventing bandwidth, latency, and cost
concerns of cloud computing with the global AI edge chipset revenue forecast to
grow to $51.9B by 2025 [24].

2.1 Background
The 5G NR air interface is vastly improved to address the requirements of
various emerging use cases. To guarantee QoS requirements for such use
cases, other network architectural concepts have been introduced that support
management and orchestration, optimization, and AI/ML-based analytics. A
new split RAN architecture and edge computing are two significant architectural
changes in 5G networks to reduce the overall latency and guarantee QoS.
With the migration to cloud-based 5G networks, there is a need for a
collaborative and intelligent approach to optimize the fragmented ecosystem
of edge computing. While mobile network operators can open up their network
as a distributed cloud to non-telco workloads, enterprises need to optimize
their applications for the new distributed architecture, which provides an
unprecedented opportunity for distribution and processing of the massive
amounts of data and its analytics at the edge. 5G networks allow us to fully
exploit edge computing by moving the data collection, compute, and analytics
closer to the end points, where data is generated and consumed, rather than
sending the data to and from servers in cloud data centers thus essentially
leading to significant reduction in end-to-end latency in data analytics and
delivery , 5G and beyond edge networks inherently have the intelligence needed
for smartly moving, storing, and processing data on the fly.
With both massive data and the processing power with AI/ML capabilities at
5G edge nodes, including mobile edge computing nodes and mobile nodes,
17 5G Edge Automation & Intelligence| November 2021

distributed learning [25] and collaborative intelligence [26] will become possible
and be able to support real-time intelligence and collaboration. Edge nodes can
work together to share sensor data from each other to obtain joint perception
with collaborative AI/ML in a dynamic environment and to act together with
group decisions for the improvement of efficiency, productivity, and safety
for various 5G edge applications like intelligent transportation system, smart
factory, smart energy, and smart homes.

2.2 State of the Art and Industry Landscape

5G calls for a new level of flexibility in architecting, scaling, and deploying
telecom networks. Cloud technology offers new innovative alternatives for such
RAN deployments complementing existing proven purpose-built solutions.
Cloud RAN refers to realizing RAN functions over a generic computing platform
instead of a purpose-built hardware platform and managing the RAN application
virtualization using cloud-native principles. Cloudification of the RAN begins with
running selected 5G RAN network functions in containers through Commercial
Off-The-Shelf (COTS) hardware platforms. It starts with the control plane and
user plane in the CU and continues with latency-sensitive radio processing
functions in the DU. By pushing distributed units to the edge, mobile networks
provide low latency services and a pool of processing and other computing
resources to support mobile user processor off-loading use cases [27].

2.2.1 The Industrial Cloud Ecosystem

The combination of different sizes of cloud data centers, namely edge zones, at
global, national, local/regional, and potentially access locations are integrated
into the network and operated by a central orchestration and management
system. The exact specification of the infrastructure on different sites may
depend on the use cases and applications onboarded. In addition, there can
be several infrastructure providers on the same site and these distributed
computing resources, including MEC nodes and edge devices like the Roadside
Unit (RSU) which can provide AI/ML services [28] or enable collaborative
intelligence.
Many companies participate in the ecosystem, from hardware vendors, platform
companies to applications developers, System Integrator (SI) companies, and
Cloud Service Providers (CSPs). Two other key players in the ecosystem are the
Hyperscale Cloud Providers (HCPs) and Operational Technology (OT) vendors.
Hyperscale Cloud Providers, such as AWS, Microsoft Azure, Google, and AliCloud
all have a core business to provide cloud infrastructure and platforms. They
own application ecosystems with thousands of contributing developers and
can serve multiple enterprises in several sectors globally. HCPs are keen to be
ecosystem drivers for edge computing. As a part of this approach, HCPs start
offering on-premises compute, storage, database, and other services run locally
on dedicated platforms provided to the customers at the edge. This approach is
ideal for workloads and applications that require low latency and access to on-
premises systems, enabling edge workloads to extend their reach to the cloud as
needed.
OT vendors have IoT platforms and applications, supported by edge computing
components. Some examples of these companies include Siemens,
General Electric, BMW, and ABB. They have strong enterprise relationships,
especially in the manufacturing sector. Companies looking to do an intelligent
manufacturing deployment are likely to partner with an OT vendor to a certain
extent. OT vendors establish relationships with HCPs for global deployments
of their solutions, access to the application development ecosystem, and an
environment to create and deploy their applications.
SI companies have a wide range of capabilities to address enterprise pain
points related to solution implementation and integrating offerings from
18 5G Edge Automation & Intelligence| November 2021

Figure 8 Distributed edge to meet the need of 5G applications.

different ecosystem companies. SI level), high reliability (up to 99.999%) 2.2.4 RAN Intelligence
companies can be both global and and high throughput (from 1Gbps up Controllers
local and are likely to be present in to 10Gbps) at the edge, which can
most solution implementations in one further enable edge applications to In a typical RAN, there are millions of
way or another. Apart from specialized provide localized, real-time, safety decisions taken every second about
SI companies, other companies related AI/ML services. which user to serve over the radio
can also take an SI role in solution interface and how. Each of these
implementation, for example, OT 2.2.3 Data Collection and decisions contributes to the service
vendors or HCPs. Analysis at Edge quality and the prioritization among
users and services in case of conflicts.
Figure 8 provides a better look at the One key benefit of the edge Traditionally, these micro-decisions
overall cloud and edge ecosystem, as intelligence is the accessibility of are governed by a combination of
it pertains to 5G applications. real-time, local data that is captured supplier design choices and network
by nearby devices or sensors configuration parameter settings done
2.2.2 Centralized, connected to the edge or mobile edge by the service provider. In the relatively
Decentralized, and Hybrid computing node. 3GPP has defined simple 2G systems, the effect of a
data collection features like [31]. In
Network at Edge configuration change was mainly
the O-RAN architecture, data can be possible to understand. In today’s
Edge intelligence requires non- collected at the edge nodes including more sophisticated multi-service 5G
traditional hierarchy, a hybrid base station, mobile edge computing networks, it is virtually impossible,
and distributed network for edge node, and user devices, to support in a cost-efficient manner, to predict
devices. The introduction of 5G near real-time network analysis the effect a given set of configuration
NR sidelink [29] and Integrated for RAN optimization. Similarly, changes will have on the end-user
Access and Backhaul (IAB) [30] in other sensor data like video, audio, services.
3GPP standards has enabled both temperature, and others could be
direct device-to-device and multi- collected by various edge nodes and However, the intent of the RAN
hop communication at the edge can be used for various edge AI/ML remains the same: to offer
of networks. They enable a hybrid applications to help provide near-real connectivity to the service providers’
network topology of both centralized time analysis and decision making, to customers in a profitable way. The
and distributed (mesh) networks at enforce safety, security or efficiency idea of intent-based management for
the edge, allowing the edge nodes to in verticals such as intelligent RAN is to evolve the RAN configuration
communicate without going through transportation systems, smart factory from setting technical parameters and
the hierarchy of the core network. cases, and others. instead allowing service providers to
Such heterogeneous connections specify the connectivity service itself,
and networking technology can thus prioritizing across users and services
enable low latency (at milli-second based on business intent and devices
capabilities.
19 5G Edge Automation & Intelligence| November 2021

The non-real-time RIC is a concept • AI/ML over-engineered Neural Networks (DNNs), which are
developed by the O-RAN Alliance to programs: The intelligence in considered complex black box models.
realize intent-based management, RAN control is gradually moving
The inability for humans to see inside
built on principles of automation and to AI/ML-based software, and the
black boxes can result in AI adoption
AI and ML. The non-real-time RIC non-real-time RIC is designed for
(and even its further development)
brings genuinely novel capabilities to AI/ML from day one.
being hindered, which is why growing
the system and addresses use cases • Innovation for openness: It is levels of autonomy, complexity, and
that were previously out of reach, possible to build an open eco- ambiguity in AI methods continues to
with the ability to set policies per user system of intelligent controller increase the need for interpretability,
and data enrichment information software where applications transparency, understandability, and
for RAN optimization. Intent-based (rAPPs) feed each other with data explainability of AI products/outputs
management based on non-real-time and insights. (such as predictions, decisions,
RIC can be applied to Cloud RAN
actions, and recommendations).
to enable a high degree of network 2.3 Envisioned Features and These elements are crucial to ensuring
programmability and can equally well Key Technologies that humans can understand and
be applied to purpose-built RAN to
Multiple international organizations — consequently — trust AI-based
enable a wide variety of automation
have defined the expected systems. Explainable AI (XAI) refers to
and optimization use-cases that are
requirements, features, and key methods and techniques that produce
not possible today. This approach can
technologies in the context of 5G edge. accurate, explainable models of why
be extended to the resources at the
ESTI has published a specification [32] and how an AI algorithm arrives at a
edge and combined with computing
on an MEC framework and a reference specific decision so that AI solution
resource allocation to the end-users.
architecture, as well as many other results can be understood by humans.
The non-real-time RIC is part of specifications as summarized in the Without explanations behind an AI
the Service Management and MEC in their 5G Network white paper model’s internal functionalities and
Orchestration system (SMO) and [33]. 3GPP has defined even more key its decisions, there is a risk that
consists of a platform plus a set of enabling technologies of 5G to support the model would not be considered
microservices (named rAPPs by O-RAN MEC [34]. In addition, concepts trustworthy or legitimate. XAI provides
Alliance) representing the network such as Explainable AI, Named the needed understandability and
intelligence. The system’s design is Data Network, innovative transport transparency to enable greater trust
based on the following principles: layer protocol, joint optimization of toward AI- based solutions. Thus, XAI is
• Access to information: There is a communication and computing, and acknowledged as a crucial feature for
wealth of contextual information distributed machine learning have the practical deployment of AI models
– not available in the RAN – been studied by various academic in systems and, more importantly, for
with the potential to improve and industry organizations. These satisfying the fundamental rights of
radio–resource management, new features and key technologies will AI users related to AI-based decision-
RAN performance, and user play an important role in future edge making (according to European
experience. This includes optimization, intelligence, and data Commission ethical guidelines for
application-level information, analytics. The following sections will trustworthy AI). Standardization bodies
cross-domain information, UE give an overview of some which may such as the ETSI and the Institute of
positions, mobility trajectories, be growing in importance. Electrical and Electronics Engineers
UE computation capabilities and (IEEE) also emphasize the importance
2.3.1 Explainable AI
external information. of XAI where AI models are deployed,
• Dynamic optimization: AI has achieved growing momentum indicating XAI’s growing importance
Traditionally, management in its application in many fields to deal in the future. AI deployers and
and orchestration have been with increased complexity, scalability, developers must comply with these
performed on the timescale of and automation, which also permeates ethical guidelines and regulations
hours. With automation and digital networks today. A rapid surge to ensure their AI solutions are
improved interfaces, the non- in the complexity and sophistication explainable and trustable.
real-time RIC can optimize the of AI-powered systems has evolved to
such an extent that humans do not However, there are significant
RAN on a time scale down to half challenges in developing explainability
a second. understand the complex mechanisms
by which AI systems work or how they methods. One of them is the trade-
• User-level service assurance: off between attaining the simplicity
make certain decisions — something
Optimizing the RAN on a user of algorithm transparency and
that is particularly a challenge when
level (in addition to the per-node impacting the high-performing nature
AI-based systems compute outputs
level) enables the non-real-time of complex but opaque models (when
that are unexpected or seemingly
RIC to address a broad set of one increases the transparency
unpredictable. This especially
use-cases that were previously aspect, privacy and the security of
holds for opaque decision-making
out of reach. sensitive data come into question).
systems, such as those using Deep
20 5G Edge Automation & Intelligence| November 2021

Yet another challenge is to identify • Confidence, which is achieved and the environment through properly
the correct information for the user, through having a model that is defined interfaces.
where different levels of knowledge robust, stable, and explainable
The main components of such a
will come into play. Beyond selecting to support human confidence in
framework center on explanations,
the level of knowledge retained by the deploying an AI model
explainability for data, explainability
user, generating a concise (simple but • Privacy awareness, ensuring for ML, and explainability for MR
meaningful) explanation also becomes that the AI and XAI methods do (see purple parts of Figure 9). The
a challenge. Researchers attempt to not expose private data (which distinctive approach that we are taking
integrate knowledge-based systems so can be done through data is to apply explainability to ML and MR
that the explanation becomes relevant anonymization) and the interplay between ML and MR
to its application’s context [35].
• Actionability, with XAI providing by feeding the output of an ML model
XAI helps deliver trust by supporting indications regarding how a user (both its predictions and explanations)
with the following properties: could change an action to yield into our MR techniques and applying
a different outcome in addition it XAI to generate explanations. This
• Trustworthiness, to attain the to providing the rationale for an proactive placement provides the right
trust of humans on the AI model outcome AI trustworthiness early on rather than
by explaining the characteristics
• Tailored (user-focused) relying on reactive fixes. Furthermore,
and rationale of the AI output
explanations, allowing humans this framework allows the integration
• Transferability, where the — as AI system users of different of new XAI algorithms into the
explanation of an AI model knowledge backgrounds — to respective explainability components.
allows a better understanding understand the behavior and In the future, newly developed XAI
of it so that it can be transferred predictions made by AI-based techniques for ML/MR can be easily
to another problem or domain/ systems through tailored deployed within the explainability for
application properly descriptions based on their roles, ML/MR components.
• Informativeness, relating to goals, and preferences
informing a user regarding 2.3.2 Multi-Access for
It is vital to incorporate interpretability
how an AI model works to and explainability at different levels the 5G Edge
avoid misconception (this is of complex AI techniques. The XAI Multi-access traffic management at
also related to human agency framework is tightly linked with the edge is vital for addressing ever
and autonomy, which ensures providing explanations for both increasing performance requirements
humans understand AI outcomes different AI techniques (ML and for current and future applications.
and can take intervening actions Machine Reasoning (MR) techniques) However, it is not possible to achieve,
on that basis)

Figure 9 Overview of XAI methods and their link to data, ML, and MR
21 5G Edge Automation & Intelligence| November 2021

for example, the maximum data rate and the highest reliability at the same time.
Hence, the multi-access challenge is to manage data traffic across all available
access networks and meet diverse application requirements in coverage, rate,
latency, and reliability. To address this challenge, the following key questions
need to be answered:
• How to acquire individual application requirements?
• How to support multi-access traffic management?
• What measurements are needed for making smart decisions?
Applications may have different QoS requirements, and Traffic Management
(TM) service [36] recently introduced in the ETSI MEC reference architecture
[37] allows applications to get informed of various capabilities and multi-access
network connection information, and to provide requirements such as delay,
throughput, and loss for influencing traffic management operations at the edge.
Multi-access traffic management requires a set of new protocols between client
and network. Recently, multiple access management service [38] has been
proposed. In parallel, 3GPP has developed the access traffic steering, switching,
and splitting [39] feature. Both provide mechanisms for flexible selection of
network paths, and leverage network intelligence and policies to dynamically
adapt traffic distribution across selected paths under changing network/link
conditions. Figure 10 shows the multi-access protocol stack which consists of
the following two sublayers:
• Convergence sublayer: This layer performs multi-access specific tasks
such as access (path) selection, multi-link (path) aggregation, splitting/
re-ordering, lossless switching, keep-alive, and probing. Generic Routing
Encapsulation (GRE) [40] may be used to encode additional control
information, e.g., sequence number, at this sub-layer.
• Adaptation sublayer: This layer performs functions to handle tunneling,
network layer security, and Network Address Translation (NAT). Existing
protocols, including User Datagram Protocol (UDP) and Internet Protocol
Security (IPSec), can be re-used.
To take full advantage of multi-access connectivity, we should distribute traffic
load intelligently across available access links in a manner that improves
user experience with efficient radio resource usage. To achieve this goal,
measurements that reflect the connectivity conditions of different access
networks should be incorporated while determining multi-access traffic
distribution. For example, the end-to-end packet delay measurements can be
used to identify which access provides better latency performance.

Figure 10 Multi-Access Protocol Stack.

22 5G Edge Automation & Intelligence| November 2021

When serving QoS flows requiring Network Slice Instances (NSIs), each conjunction with the application
high reliability, packet drop ratio implementing the same slice but for awareness for building a stateful way
measurements give a good indication different geographic locations, will be of serving dynamic and situational-
of whether redundant transmission important here. aware network.
over multiple access networks is
required. In addition to end-to-end 2.3.4 Situation-aware 2.3.5 Joint Optimization of
packet statistics, RAN measurements Transport Layer Protocol Communication and Computing
can indicate network quality
Transport layer protocols such With the proliferation of smart
degradation caused by deteriorating
as Transmission Control Protocol autonomous devices equipped
radio link quality or access network
(TCP) rely on the end host to handle with computing, communication,
congestion, in a timely fashion.
congestion control, flow control and storage, we are observing data
Moreover, ML/AI based algorithms
and end-to-end reliability with the explosion at the edge, generated
may be developed to automatically
underlying assumption of ‘in-order from many new applications such as
configure and manage data traffic
byte’ delivery. However, dynamic, self-driving cars, AR/VR, panoramic
across all the available networks and
self-organized situational 5G networks telephony, holographic interaction,
improve end-user experiences.
at the edge, which may be running digital twins, etc. 5G edge computing
Using these building blocks, we can diverse applications (from end users is essentially bringing computing,
unleash the full potential of multi- to edge) in a collaborative, multi- intelligence, and communication
access/multi-connectivity at the hop fashion with involvement of together by providing a platform to
edge to address the performance multiple edge infrastructures, may process data close to source for
requirements of applications now and not necessarily stick to the in-order these services, thereby reducing
in the future. byte delivery paradigm as the network end-to-end latency and network traffic
topology itself may be dynamically overhead. Both communication and
2.3.3 Situational Network at changing [41], [42]. edge computing resources, however,
the Edge are limited, variable, and shared by
On the other hand, TCP relies on its several nodes in the vicinity running
The emerging architecture being reactive congestion control which multiple applications with different
designed to accommodate 5G kicks in only after the packet has performance requirements.
applications is known as Multi-Access traversed full path from the source
Edge Computing (MEC) and several node to the destination and back to The existing computing and
challenges must be solved before it the source/host edge infrastructure. communication layers have mostly
can become a mainstream reality. Such a reactive approach leads to been designed either in an isolated
delays. Furthermore, the traditional manner or via loose coupling with
First, devices that might need edge end-host networking-based approach coarse-grained information sharing
computing must be recognized at the may be prone to single point of failure between the layers. The expectation
beginning of a session when user (or attacks by external jammers or with 5G is that it will increase both
plane resources are being established.
misbehavior by existing end-users/ users and traffic by an order of
A local uplink classifier or branching subscribers of the network services) magnitude [43]. For this reason,
point or packet session anchor which may overload the networks, thus it is of paramount importance to
must be allocated for these devices. resulting in highly unreliable and latent jointly optimize the computing and
Recognition of such devices may not networks at the edge. Thus, the added communication process to help avoid
be straightforward and could present dynamics due to mobility, changing underutilization and oversubscription
an interesting machine learning deployment and reconfiguring topology of resources, while also significantly
challenge in and of itself. Second, need to be taken into consideration improving an application’s
protocols and techniques must be while designing Transport Layer performance.
created for devices to discover any Protocols.
edge services available to them, While the concept of computing-
including discovery of peer devices To this end, there is an inherent communication co-optimization
that may also be attached to the need for design of a situation-aware has been under consideration by
network as clients. Third, mobility transport protocol. Additionally, academia and industry for a while,
management must be developed for some awareness of the network there are several challenges and
the unique edge environment that can be brought into consideration directions that require closer attention
preserves the IP address associated while designing such Transport to unleash the full potential of the joint
with a given session when necessary Layer Protocols. For instance, one optimization. First, a sound theoretical
but that also recognizes that IP such protocol may involve in-band, framework is indispensable to study
addresses may need to be reassigned telemetry-based network awareness the optimal joint resource allocation
under some circumstances. Finally, building at the transport layer by, for and resulting performance benefit
the interactions between edge example, exposing the congestion- for a given topology, Radio Access
computing and slicing must be related meta data. Additionally, Technology (RAT), and computing
dealt with. The concept of multiple this mechanism can be used in resource (trusted peers, edge, cloud).
23 5G Edge Automation & Intelligence| November 2021

The framework should consider to develop a realistic simulator/ leveraging disaggregated resources for
realistic modeling of communication emulator leveraging open-source edge compute and processing.
and computing processes. computing frameworks (e.g., O-RAN,
Distributed learning approaches
OpenNESS), AI/ML frameworks (e.g.,
Second, it is important to have an have evolved to address several
Intel® Distribution of OpenVINO™
in-depth characterization study of issues of centralized cloud-based
toolkit), AI/ML libraries, and real-
the emerging edge workload and learning and are especially relevant
world sensor datasets, which will be
design a generic computing resource to the problems at the 5G edge.
handy to validate the scalability of
representation or abstraction to In particular, distributed learning
the aspects mentioned above and to
assess the computing storage uses are inherently collaborative at
generate key performance indicators.
requirements. While AI/ML is likely to typical endpoint or edge nodes which
be the dominant workload to enable The joint optimization problem poses have only a partial view of the data
edge intelligence, the computing and additional challenges and complexity required for learning. Collaboration in
communication resource requirements in the case of edge-enabled distributed learning may take different
for distributed training (e.g., federated applications such as control systems forms, such as in Federated Learning
learning) and inference are different. and robotic applications, where or Fully Distributed (Decentralized)
The network also needs to collect time-sensitive functions are offloaded Learning, [46]. In Federated Learning,
real-time status or telemetry on the from robots to the edge for compute the collaboration is managed with
usage of current computing and acceleration, energy efficiency and for the help of a central coordinator that
communication resources in an leveraging advanced AI capabilities. combines the learnings from nodes
efficient manner. Guaranteed latency and determinism processing over their own data.
are critical for reliable control loop Whereas in Fully Distributed Learning,
Third, the applications are increasingly operation, especially while off-loading there is no central coordinator, and
becoming distributed and being robotic functions (such as perception, nodes must collaborate in a peer-to-
implemented using cloud-native planning, cognition) on the edge at peer manner.
constructs. Additionally, applications scale. As a result, robotic control
can be decomposed dynamically An additional advantage of distributed
also needs to be jointly optimized
(such as distributed inference [44]) learning that is relevant to the 5G
along with communications and
where the real-time availability of edge is the ability to adapt the
computing for reliable and efficient
computing and communication learning models to the local situational
robotic operation. The state of the
resources will help determine the right context of the edge. For example,
wireless network (latency, packet
decomposition and placement/off- a 5G edge infrastructure for a
errors etc.) can be used to adapt
loading of decomposed units. factory implementation would have
robotic control to changing network
different requirements compared
Fourth, the traditional design conditions and available compute
to a 5G edge network deployed to
of scheduling computing and resources. Similarly, the state of
service urban residential clients.
communication resources should robots and their environment can be
In this case, a distributed learning
be re-thought separately. Depending used to dynamically adapt, provision
framework deployed to optimize the
on the underlying hardware (e.g., and schedule compute resources
5G edge for functions such as traffic
Central Processing Unit (CPU), (CPU cores, memory, etc.) and
cell prediction, QoS management
Graphical Processing Unit (GPU), communications resources (packet
may need to be adapted differently
Field Programmable Gate Arrays scheduling, reliability) to the changing
for the two distinct deployments.
(FPGA), custom accelerators) used needs of robotic tasks.
Distributed ML allows for such model
for computing and location (e.g.,
2.3.6 Distributed personalization to these distinct
on-device or edge) of computing, deployments, while still benefiting
the execution time may vary and put Learning at Edge from learning common features
variable budget for the communication Distributed Learning [46] is one of through collaboration across all the
process to finish and meet end-to- the key enablers of edge intelligence nodes. With distributed learning local
end application requirement. A joint that focuses on both training and context or situational awareness can
computing-communication resource inference with private and sensitive easily be included within the locally
scheduler can leverage the real-time data at the edge, while avoiding the trained ML models to improve model
information [45], use AI/ML-based communication and latency cost of accuracy for the local 5G edge context,
techniques to predict the arrival of moving data for centralized processing when compared to a centrally trained
input data for computing, and finally in the datacenter. At the same time, model.
schedule both resources to keep the it enables efficient use of compute
computing instances (e.g., container) capabilities available at the network Distributed learning has many
up and running for processing data edge or across a group of on-premises challenges, particularly when applied
upon arrival. devices pooled together. Either of such over the 5G wireless edge [46]. We list
available compute capabilities can some of the challenges below:
Last but not the least, academia
and industry can work together thus host services by harnessing data
close to the generation points and by
24 5G Edge Automation & Intelligence| November 2021

• Statistical properties of data are one of the key compute workloads will be configured by the application
distribution: Different nodes that must be supported over the developer. The application running
at the 5G edge may have 5G Edge. Hence, several compute- on the network edge may have
different data collection and communication co-optimization connectivity requirements on
storage abilities, as well as may approaches discussed in earlier bandwidth, throughput, mobility, and/
only have a limited view of the section are also relevant. or latency within its components
overall data distribution. The (for example, deployed on different
Currently there are various efforts to
diverse non-Independent and hardware in a redundant setup) or with
apply distributed learning solutions
Identically Distributed (non-IID) the external world, such as an internet
to applications in wireless networking
data across clients can lead to connection, and the user equipment
[50] [49]. While there are recent
slow convergence of ML model or the application session. Traffic
efforts to introduce distributed
training slow down AI model routing for applications deployed
learning techniques in both O-RAN
training. further out in the network topology will
and 3GPP, such efforts are nascent.
• Heterogeneous Communication need new mobile network solutions,
Further work will be required to evolve
and Computational Costs: such as distributed anchor, session
5G edge architecture framework to
Devices on the wireless edge breakout, and multiple sessions, and
support distributed learning as a
may have diverse computational in some cases coordination between
critical workload, in the design of the
and communications capabilities application server selection and usage
next generation of wireless standards.
causing ‘straggler effects’ where of these mobile network solutions.
poorly resourced clients on Application runtime execution
inefficient links relay their data 2.4 Requirement Analysis
environment: The very basic
infrequently or have high error The ecosystem for edge computing is functionality that an edge computing
rates. This leads to problems with fragmented and is quickly evolving. service may provide is the runtime
model convergence, accuracy Technical solutions, interfaces, execution environment for VNF and
and fairness. standards, and business models non-telco workloads. An execution
• Scalability: Collaboration to learn are not set. Several players must be environment should be able to host
very large models across a large involved to create end-to-end solutions applications and harmonize the
number of users can lead to poor and CSPs must carefully consider in requirements of the development
scaling. which industries they can expand their communities. Many applications may
offerings beyond connectivity. use edge computing with different
• Privacy and Security: While
distributed learning solutions The edge application ecosystem is characteristics and functional
avoid data sharing across driven by third-party applications requirements and require different
devices, sharing of model outside of the telecom domain platform components. Therefore,
parameters during learning can since solutions for new use cases the operator provides a generic or
still leak data privacy. Studies require specific domain knowledge multiple execution environments on
have shown that it is possible to from industry players outside the the network edge that application
reconstruct user’s data through telecom space. Edge infrastructure developers can later customize.
model inversion attacks. will therefore be accessible to third Dynamic orchestration and
• Security risks: Adversarial or party application providers and management: Centralized
malicious clients can corrupt developers and will host a multitude orchestration is required to maintain
model training by inserting false of applications, each with specific consistency between possible traffic
updates. characteristics and needs. breakout points (where user plane
• Need for Self-Learning. Another The edge application environment gateway functionality is deployed)
challenge involves learning with enables mobile network operators to and the applications (which consume
limited data labeling, as access host non-telco workloads and open the traffic) in the network edge.
to annotated data with limited up the network as a distributed cloud The central orchestration and
human support is one of the resource. Enterprises can develop management functionalities need
important challenges for the applications, deploy, and manage to be aware of the network topology
edge environment. them flexibly via orchestration logic and the available resources in the
• Continuous learning: Constant towards a ‘landing zone’ that accesses distributed cloud infrastructure.
new updates to the model may the distributed cloud infrastructure This orchestration layer will provide
lead to a catastrophic forgetting and leverages services exposed a harmonized single orchestration
of the model’s earlier learning. through APIs for consumption. Below and management functionality over
is a brief overview of the functional the different orchestration functions
There are several promising
components needed to create an edge present. One of its purposes is to
techniques that are being explored to
computing solution [51]. manage the platforms for non-telco
mitigate the above issues and may workloads and VNFs according to
be found in [46] [47] [48] [49]. We Connectivity: Once the development service level agreements.
further note that ML computations environment is installed, connectivity
25 5G Edge Automation & Intelligence| November 2021

Service exposure: Exposure is a crucial architectural innovations are required intelligent collaboration among
function of defining and developing to support the envisioned features sensors, infrastructure nodes, and
new capabilities (APIs) and securely and key technologies at 5G edge. local compute nodes to process data
exposing them to non-telco workloads. First, at edge network level, traditional closer to its source or point of service
The exposure server exposes the 5G connectivity provides centralized, delivery.
core capabilities available internally hierarchical architecture plus limited
Situational awareness can be acquired
within the operator or to a partner device-to-device connectivity support.
and maintained by mainly two types of
with a commercial agreement. The Due to the high dynamic nature of
systems: (i) a dynamic context-based
exposed core capabilities add value to the edge, it is needed to have highly
discovery system, (ii) an Intelligence,
internal or external users, for example, flexible situational networks at the
Surveillance, and Reconnaissance
connectivity, optimization, identity, edge to connect the edge devices
(ISR) system. In dynamic context
security, data, and analytics. on the move to edge nodes, and
discovery, each node in situational
connect the edge devices to each
Optimization: 5G and edge computing edge network continuously acquires
other, with or without centralized core
techniques provide several and maintains updated situation
networks support. Such situational
opportunities for smart network perception and network context
network at the edge will allow
optimization, which can be theoretical, in the proximity and network by
sharing of data, sensor resource,
heuristic, or AI/ML-based. AI/ML frequently sharing information such
and computing processing capability
techniques can 1) detect changes in as environment perception, Node’s
with low latency, high reliability, and
demand, deterioration or drifting of own status and information (device
high flexibility at the edge for various
SLAs, and inefficiencies or problems in type/role, location, orientation,
use cases. Secondly, to support the
the network, 2) diagnose such issues etc.), perceived communication
ever-increasing AI/ML computing
and identify the root cause, and 3) environment, compute capability, and
needs and exponentially rise in AI/
predict the response of the network to sensing capability and configuration.
ML workloads, distributed learning
workload redistribution, deployment of Context discovery enables the edge
at the edge is needed to leverage
new resources (e.g., network slices), situation network to form an intelligent
the distributed data as well as AI/
configuration changes, and changes collaborative group for efficient and
ML processing capability of scattered
in management policies. The system intelligent optimization of sensing,
edge network nodes and edge devices.
can then select and implement the caching, communication, and compute
This calls for a new distributed AI/
best response to changing conditions. requirements in the network. Context
ML architecture to make optimal
Operators can use AI/ML techniques information maintained by dynamic
use of the available communication
to gain useful and timely insight context discovery is then utilized to
and computing resources while
into their networks and optimize the create a collaborative ISR system
meeting the latency, security, and
management, operation, and/or by pushing compute at the edge to
privacy requirements the 5G edge.
orchestration of the network. realize data to decision concept in a
The following sections discuss these
time-sensitive way. ISR optimizes the
To realize these functionalities at the architecture directions in detail.
utilization of sensor assets, compute
edge, there is paramount need for
2.5.1 Situational resources and network resources
building aggregate yet distributed to collect and fuse actionable
knowledge of the applications, Network Architecture information to provide reliable high-
situations, or workloads, across all the Edge computing provides an ideal quality situational information for
devices and entities involved in such platform to enable many critical assessing options, threats, and
movement, storage and/or processing and time-sensitive applications that consequences of decisions. A
of the data at the edge. Inherently, require huge sensor capabilities situational aware model also needs to
such knowledge building would require and computing resources to provide meaningful representations
intelligence and judicious collaboration process sensor data in near real- of actionable context and situational
between all the network entities/ time. Furthermore, intelligent data information so that network and users
devices involved thus requiring the movement in a bandwidth-efficient can readily consume the information
need for collaborative intelligence at manner and data utilization to make in optimizing various operations and
the edge. Towards this end, a few key intelligent and timely decisions by time-critical decision making.
innovations in the areas of distributed running AI/ML algorithms at the edge
learning, edge data analysis for driving must be accommodated at the edge. 2.5.2 Collaborative
scalable intelligence across multiple Situational awareness becomes Edge Intelligence
applications/workloads with bounded critical here to take maximum
latency and guaranteed high reliability Edge intelligence, where intelligent
advantage of potentially big data
are needed. compute devices are needed for
generated by sensor system utilizing
moving, storing, and processing data
compute resources available at edge
2.5 Architecture Direction closer to its source or point of service
in a time-critical way while keeping the
delivery is paramount for prediction,
To realize the 5G edge optimization, bandwidth requirements manageable.
preparation, and response in an
edge intelligence, and data analytics, An edge situation network enables
accelerated manner to deliver near
26 5G Edge Automation & Intelligence| November 2021

real-time services. Edge intelligence is event detection, action classification, where low latency and high bandwidth
paramount for real-time services and and collaborative decision making networks allow for many edge devices
can accelerate AI/ML computations with a comprehensive application- to perform numerous communications
and workloads and offload centralized and-context or situation awareness of over a constantly growing variety of
systems (e.g., cloud-based) that the edge environment in which such applications. ML-driven optimization
require higher bandwidths and intelligent agents operate. has the potential to adapt to
lower latencies. Many applications evolving situations, conditions and
The CEI provides a paradigm for
leveraging AI/ML at the edge, however, heterogeneous environments as well
structured collaboration among the
are real-time and collaborative in as seeing through complex interaction
intelligent edge agents so that the joint
nature where the form of collaboration between the various components of a
edge intelligence can be realized to
can be fluid and application-and- system and optimize resource usage
attain an overall objective of delivering
context dependent (e.g., local in a way that is rarely accessible, even
real-time response service for, say,
relevance). For example, multi-camera with domain expertise.
data and analytics delivery to edge
video analytics at smart intersections
users. A fundamental enabler for CEI 2.6.1 ML for Systems
requires real-time communication
is intelligent networked computing to
with the smart AI cameras (that have There are many ways in which ML can
enable near real-time collaboration
the most relevant field of view) for be used to optimize 5G. Data delivery
among heterogeneous computation
sharing information such as location, through the network may be required
capable edge servers and edge
frame, output of local processing and to be lossless (e.g., state-dependent
users. Hence, networked computing
analytics. encryption) or can sustain losses
framework, algorithms, application-
To enable such near real-time aware communication-compute (e.g., audio or video streaming). For
application-and-context or situation- protocol built on top of ultra-low lossless communications, it is desired
aware collaboration among the latency, guaranteed highly reliable to have intelligent control over the
intelligent edge devices and edge (and always available) communication way data is processed and paced
infrastructures, the underlying 5G substrate, are fundamental enablers through the network and, finally,
networks needs to offer ultra-low to realize CEI. delivered to its destination. For lossy
latency and ultra-highly reliable communications, it may be smartly
For realizing such CEI, the diverse controlling the loss (e.g., lower the
communication. When these
workloads across AI, media, and resolution of a video stream upon
intelligent edge users are connected
network, all together converge onto congestion).
over a 5G network they can, together
a common infrastructure which must
with the edge infrastructure, create
deliver optimization, efficiency, and 2.6.1.1 ML for Congestion Control
a locally-available shared and
lower cost of ownership. With modern Congestion control refers to a
distributed computing substrate that
state-of-the-art robust packages mechanism that determines the
can be leveraged for time-sensitive
and tools, telecommunications, pace at which a sender injects new
collective computing or analytics
semiconductors, and other ecosystem data into the network. The traditional
tasks that may be comprised of a
partners, would require the congestion control mechanism,
chain of AI/ML inferences. Such
development of such converged edge employed over TCP, relies on an
collaboration among intelligent nodes,
applications with AI and 5G networking intuitive mechanism and possibly
if done intelligently and judiciously,
capabilities [52]. with theoretical guarantees, such
can enable faster and more accurate
decision-making processes for several as “Additive Increase/Multiplicative
high-stake applications such as traffic 2.6 System Decrease” (AIMD), along with “slow
management, emergency response, Recommendations for ML- start” and “congestion avoidance.”
drones, AR/VR, and autonomous More recent mechanisms introduce
driven Optimization
systems comprising of connected rate-based methods (e.g., CUBIC),
autonomous robots, connected The joint complexity of different incorporate feedback from network
autonomous vehicles, among other edge devices, network components, switches (e.g., DCTCP), and even work
systems. communication protocols, cross layers (e.g., QUICK).
mechanisms over the different layers
To this end, the need for building and applications create a reality of The quality of the congestion control
collective knowledge and sharing very complex interactions and mutual is well-known to have a critical impact
such knowledge by forming networks influence. Manual optimization or on both network throughput (e.g., due
of collaborative intelligent nodes optimization based on classical to packet losses and retransmissions)
is paramount. Such collective algorithms and approaches requires and latency. Thus, replacing such
intelligence with connected network ever-growing domain expertise and classic solutions with ML-driven
of intelligent agents can be termed human resources. algorithms holds the potential to lower
as Collaborative Edge Intelligence latency and increase throughput by
(CEI). The emergence of CEI can An ML-driven optimization is thus having the mechanisms themselves
thus address diverse application an appealing tool for such modern adapt to evolving network conditions
requirements, for instance, real-time systems. This is the case for 5G, and considering the traffic itself.
27 5G Edge Automation & Intelligence| November 2021

2.6.1.2 ML for Data Streaming 2.6.1.4 ML for Scheduling well, and for some others it may
and Load balancing be more challenging. It is therefore
For video, audio, and gaming
of interest to expose API to the
applications, it is often the case Scheduling and load balancing are ML applications that will guide the
that latency is of greater importance at the heart of the management of prefetch mechanisms towards which
than throughput. Smart ML-driven any distributed system. The main goal data to prepare. Note that we have
data streaming aims to adjust the of these algorithms is to distribute earlier discussed ML-based caching,
signal’s quality (e.g., video resolution) and time the work in a way that however, we may be also interested
to achieve better client experience optimizes some target metrics such in classic data structure with
and resource usage. The advantage as job completion times, tail latency, significantly faster processing speed
of ML-based methods over classical or maximum oversubscription of a that only exposes parameters to ML
solutions is in resolving when processing or a network element. applications.
and where to introduce the loss
dynamically. For example, in audio It is often the case that these
2.6.2.2 Systems Support for
and video conferences, ML can help problems are computationally
in introducing loss in less important intractable, so modern solutions Federated Learning
data items (e.g., silent moments or heavily rely on heuristics. Moreover, In a federated learning procedure,
video fragments with homogeneous the metrics to optimize are usually many edge devices participate in a
background). The network can use an not sufficiently simple and the construction of an ML model. Usually,
ML API exposed by edge devices and systems include different participants the process involves a coordinator and
network stations that can guide it to a with different optimization goals a parameter server (which may be a
better use of data loss. (e.g., throughput vs. delay-sensitive single entity, centralized or distributed)
applications). Thus, even designing that coordinates a training procedure.
2.6.1.3 ML for Caching simple heuristics becomes highly At each training round, the coordinator
challenging and requires domain picks a subset of available devices
Caching is a well-known mechanism to
expertise across many layers. ML- that in turn derive parameter updates
improve data locality, resource usage,
based scheduling and load balancing based on their local data and send
client experience and dramatically
solutions have the potential to capture their updates to the parameter server.
reducing latency and applications
complex structures and dependencies The parameter server processes all
response time. Traditional cache
among metrics and participants and updates and computes updated model
management mechanisms rely on
offer better performance and with parameters.
intuitive methods and often with
limited domain expertise in each
theoretical guarantees (e.g., begin A main idea behind federated learning
specific application.
online competitive). Such methods is to protect the privacy of participants’
include the Least-Recently-Used 2.6.2 Systems for ML data. Federated learning imposes
(LRU) policies, their approximations, several challenges in privacy as well as
and related more recent variants Different layers of the infrastructure
bandwidth and compute, especially for
(e.g., Time-aware LRU (TLRU), Least- (i.e., network and compute) offer
edge devices. 5G has the potential to
Frequently-Used (LFU)). None of them different abstraction levels. Usually,
take federated learning a step forward
is optimal for specific application and it is the case that support of APIs
by providing support for better privacy
usage patterns. Moreover, all these exposed by the infrastructure can
(e.g., differential privacy, secure
policies are reactive because they rely offer better resource efficiency.
aggregation), support for in-network
only on previous data usage patterns. For example, while it is possible to
processing of updates (e.g., filtering,
implement lossless packet delivery
ML-based cache management may averaging, shared-randomness) and
in the application layer (e.g., over
enable two significant advances that incremental computation alleviating
UDP) the efficiency and the usage of
can lead to improved resource usage. the burden on the recourse-limited
network resources is better managed
First, the policy can be made adaptive, edge devices.
where such delivery is supported by
evolving, and learning as data keeps the network itself. Therefore, having
arriving. Second, it may learn not 2.6.2.3 Systems Support f
the 5G infrastructure to support and
only the data patterns but also take expose APIs to ML application is a key or ML SLAs
into consideration other events. For step towards resource efficiency. Different ML models trained towards
example, it may be absolutely time- the same main optimization goal may
dependent and react to real-time 2.6.2.1 Systems Support for differ in properties such as expected
events, such as parsing news sites Caching ML Application Data accuracy and inference time. To
and social media to predict what data adhere to SLAs and provide better
will be consumed shortly and prepare Data prefetch is a well-known
user experience, the infrastructure can
for it. technique to accelerate response
direct queries to those models that
times and increase resource efficiency
would best benefit the users’ needs
both in hardware and software. For
and SLAs. For example, it may decide
some applications the access patterns
to direct a query towards a faster
are evident and data pre-fetch works
28 5G Edge Automation & Intelligence| November 2021

model due to network congestion to reduce the response time.

2.6.2.4 Systems Support for Collaborative Intelligence

One of the exiting directions for nowadays and future ML applications is
Collaborative Intelligence where the edge devices communicate and interact
directly for ML purposes. One example for such interaction would be smart cars
that communicate and exchange information about traffic conditions and fast
evolving situations to prevent life threatening events.
System support for collaborative intelligence may offer short communication
paths, privacy, security, and authenticity of data and even help in processing as
discussed earlier.
29 5G Edge Automation & Intelligence| November 2021

3. Application of 5G Edge
Automation and Edge Intelligence

With the enablement of automation, optimization, and intelligent decision-

making for network and compute resource allocation, network function
selection, as well as workload optimization at the 5G edge, various use cases
can be realized with guaranteed QoS. This chapter provides a brief list of use
cases together with their main challenges and describes how they can benefit
from the application of automation and intelligence at the 5G edge.

3.1 Autonomous Industrial Solutions

Industry 4.0 and new autonomous industrial use cases for intelligent factories
bring strict requirements for computing and networking domains, challenging
broadly utilized cellular technologies. Connectivity requirements will play a vital
role in such autonomous systems. Future networks will need to support more
than 37 billion connected [53] industrial IoT devices, from sensors through
operational analytics to high-definition video analytics. From the autonomous
industry connectivity perspective, six requirements will play a critical role to
enable envisioned use cases:
• Latency – where worst-case latency requirements will drive manufacturing
automation, electrical grind and intelligent homes use cases
• High-speed bandwidth – to stream hundreds of video streams and run
distributed real-time analytics is the critical for the manufacturing to
process data
• Multi-tenancy –to provide secure access to the factory for multiple vendors
at the same time
• Autonomy / Self-configurability – for instance, where a network can self-
monitor and configure autonomously without human intervention, such as
when a freshly discovered robot can join a network on-demand and get the
required bandwidth and latency performance. Another case could involve
the autonomous re-configuring of a failed node or broken segment of a
network link.
• Security – where each part of the system, from the control plane to the
user data, must be secure and preserve privacy at each stage of data
processing
With the delivery of technologies to address the above requirements,
autonomous industrial systems will bring a new generation of tools and
technologies to increase productivity, efficiency, and safety. Tools like computer
vision, machine learning, combined with sophisticated sensing capabilities, will
improve the productivity, the inventory tracking, and the safety monitoring.
Challenges and potential approaches
While some data transfer challenges for autonomous systems can be addressed
by deploying new 5G equipment, the latency and bandwidth for industrial
applications will need to be managed through a new form of autonomous
orchestration fitted for industrial sectors. In addition to software frameworks,
new deployments will need to move away from the standard manufacturing
pyramid, to provide a unified Internet Technology (IT) and OT solutions through a
Cyber-Physical Systems overlay. Other challenges and approaches include:
30 5G Edge Automation & Intelligence| November 2021

Information Centric Networking - A The confluence of the mentioned key role in orchestrating and managing
machine-to-machine protocol like industrial edge, OPC UA, TSN, the ITS applications in a world of
Open Platform Communications industrial cloud (‘on prem’), and distributed edge computing [59].
Unified Architecture (OPC UA) [54] industry 4.0 use cases lead to Thus, it is pivotal to create common
will play a key role in data distribution significant architectural changes platforms and architectures to help
in industrial systems. However, the in Industrial Control Systems (ICS) cities merge their IoT and networking
overhead of TCP protocol when that will directly influence 5G edge workloads to achieve greater synergy
mapping names to IP addresses automation. On the other hand, the and optimize their hardware solutions
and its weak support for multicast next iteration of industrial systems for a world in which expanding cities
communication are pushing academia that will utilize 5G edge intelligence creates strain on transportation
and industrial partners to look will be able to move away from a networks. To enhance the safety,
into other approaches. One of the typical industrial multi-level factory reliability, efficiency, customer
proposed solutions is Information- overlay [56] into the autonomous and experience, and quality of a city’s
Centric Networking (ICN) [55], distributed factory automation system. transportation infrastructure, ITS
where a unified control and data technologies bring forth unique
plane can simplify the onboarding, 3.2 Intelligent Transport requirements in the form of confluence
addressing, and communication of computation and communication
patterns of industrial devices.
Systems at the edge, enabling critical services
ICN can also address increased Transportation systems are struggling for the infrastructure, vehicles, and
dynamicity challenges where drones, to keep pace with the demands of other users of the transportation
collaboration robots (“cobots”), and our globally connected economy. The system. In this way, roadside edge
mobile robots will need to move across increasing trend toward urbanization computing infrastructure forms the
the factory floor. is creating unprecedented challenges basis for realizing the ITS vision of the
for city leaders around transportation future—Intelligent, connected roadway
Time Sensitive Networking - IEEE infrastructure. In addition, today’s infrastructure that is resilient and can
802.1 Time-Sensitive Networking cities account for 71 to 76 percent adapt to the needs of a growing and
(TSN) standards and integration with of CO2 emissions [57], 67 to 76 changing city.
5G systems defined in 3GPP Rel. 16 percent of global energy use, road
will enable TSN time synchronization safety challenges with 1.3 million To this end, 5G connectivity
and scheduled traffic to be delivered deaths annually [58], and an (Vehicle-to-Infrastructure (V2I),
between TSN devices connected economic impact of $305 billion due Infrastructure-to-Pedestrian (I2P),
across a 5G system. Therefore, to to congestion. As a result, rapidly Private wireless, etc.) combined with
provide effective mechanisms to growing cities are under pressure to compute demanding capabilities
schedule traffic between TSN capable address pedestrian safety, congestion, such as multimodal sensing,
networks and 5G will be an important environmental issues, and resulting accurate positioning/localization,
capability for services such as cobots economic impact. are to be deployed at the roadside
that will autonomously work side-by- edge infrastructure to meet the
side with human workers or other Smart Roadways and Intelligent bandwidth and latency requirements
cobots. Intelligent Edge co-optimized Transportation Systems (ITS) are one for transportation. Such capabilities
with wireless TSN can also enable of the most practical and near-future would help to improve the efficiency,
compute and battery constrained applications of edge automation and safety, and experience of all the road
collaborative mobile robots, to edge intelligence. These technologies users, as well as generating an overall
effectively leverage the Edge for help enable smart cities to overcome net positive impact on environmental
advanced AI (perception, learning, some of the most pressing operational greenhouse gas emissions. The V2I
cognition and adaptation), while challenges, such as increasing and I2P communications enable
meeting tight end-to-end latency and urbanization, energy efficiency, and vehicles and other road users to
reliability requirements. congestion, that impact the daily lives communicate with static or movable
of each citizen. edge automation and road infrastructure nodes, sharing
Co-system Determinism - Finally, co- edge intelligence serve as the lifeblood data that can improve operational
systems determinism for network and for realizing an intelligent roadways coordination and efficiency. In ITS
compute where subsystems will need vision through the deployment of solutions, RSUs, vehicles, and other
to meet the expected delivery deadline multimodal sensors and real-time road users generate large amounts
every time will create new challenges processing. Together, these combine of time-sensitive data to be used
as data will need to be transported to create efficient traffic management for a variety of applications and use
often over long distances, potentially services, data collection, and real- cases. Furthermore, as the network
traversing multiple operators. New time analytics for road users and spectrum resources available for the
determinism requirements will need to pedestrian safety. ITS is limited, the data between the
be also addressed to deliver optimal nodes need to be shared in a timely
solutions for factory multitenant setup. Edge computing and automation
and reliable manner via bandwidth-
software frameworks, together with
efficient communication.
smart edge infrastructure may play a
31 5G Edge Automation & Intelligence| November 2021

Some of the important use case A Smart Home helps conserve the rich data is useful for AI/ML in the
examples for ITS include, but are not energy (electricity, gas and water), Edge, which can in turn help home
limited to: limit peak demand and increase automation. Moreover, as the edge is
overall delivery as well as endpoint closer to the home user than a public
• Sharing of perception, maneuver, efficiency. Home automation uses cloud, it reduces communication
and AI-workload models computers or smart phones to control latency. Edge may also help relieve
efficiently among entities using basic home functions and features the security concern with closed loops
the road automatically, as well as allowing between edge compute servers and
• Enhancing Vulnerable Road User vital home functions to be controlled UE. In addition, with edge automation,
(VRU) safety remotely from anywhere in the world the Smart Home can manage
• Offering value-added services through the Internet. A Smart Home resources better, for example, to
• Roadside virtual environments should provide comfort, security, monitor battery power, and utilize the
via digital twins with on-demand and the most cost-effective use of data in AI/ML training for predictive
service orchestration at the edge electricity, gas and water. Some home maintenance.
infrastructure automation can include scheduling
To better improve the home
and automatic operation of water
Addressing the unique challenges automation efficiency, AI at the edge
sprinkling, setting rules, swimming
posed by such use cases is key can be implemented partially in edge
pool conditioning, heating and air
to offering reliability, safety, and devices and partially in the hubs
conditioning, window coverings,
efficiency in transportation. Thus, and gateways through which they
security systems, lighting, food
there is immense ongoing industrial connect. With the joint AI capability,
preparation, clothes washing and
efforts for enabling ITS with active decision is more local to the user,
drying appliances, electrical vehicle
collaboration with policy makers, without sending inquiry/query to and
charging – and many more tasks.
automakers, manufacturers, cellular waiting for decision from an edge node
infrastructure operators (supporting To achieve these, the automation or a device which is far away. The
C-V2X, 5G NR V2X, Dedicated Short- system must have access to several following are some examples where
Range Communications (DSRC), sets of data, which may include: total the performance can be influenced
NextGen Wi-Fi, and beyond 5G real time energy consumption, total or improved by joint edge automation
networks) around the world. For accumulated energy usage, individual with device/hub/gateway.
seamless and unified integration of smart appliance and major load real
• Self-healing
all such technologies in common time energy consumption, pricing
socio-economic manner, it is crucial information, customer preferences • Monitor signal strength and
to actively advance the standards in terms of comfort parameters and prevent outage
and technical bodies with the cost containment, usage patterns, • Detect jamming attack on the
singular goal of using the benefits of and many more. Sources of home meters
technology across edge automation automation data can be from several • Automatic door opening
and intelligence to improve the lives sources, including: the utility company
Beyond edge automation, edge can
of citizens across smart cities and and its billing systems, smart meters
also help optimize the operations of
transportation systems. for real time and accumulated energy
components of Smart Energy and
consumption and energy quality, as
Smart Homes. As Smart Energy is
3.3 Smart Energy and Smart well as smart appliances such as
distributed to homes from central
thermostats, electric vehicle charging
Homes stations and smart sensors. locations such as factories or other
Smart energy involves electricity, energy generation facilities, it can be
water, and gas delivered to customers Additionally, there are also challenges monitored for usage at the residence,
through smart meters, which provide for home automation. As more devices thus allowing the whole energy
critical data to maximize the value are included in a Smart Home, more distribution system to be optimized
of home automation systems and data is transmitted over the network, via load forecasting, distribution
related IoT devices. A Smart Home which may result in latency concerns automation, and other energy grid
is the integration of the utility smart if the data is transmitted to the cloud optimization techniques.
meters and in-home devices enabled or the remote server is far away.
by an internal wireless radio link Another concern may involve security 3.4 Connected Health
embedded in the smart meter. There and privacy risk when home data is
transmitted to a public cloud. Accelerated by the global COVID-19
are consumer level new services and pandemic, telehealth solutions range
products such as smart appliances, Edge computing should help home from simple video conferencing
communicating thermostats, Heating, automation in multiple ways. The edge sessions with medical professionals
Ventilation, and Air Conditioning has powerful computing resources, to sophisticated in-home monitoring,
(HVAC) vent zone controllers, remote which off-loads the computing task allowing healthcare workers to track
smart phone monitoring applications, from homes. Edge networks can store progress and adjust treatments for
sprinklers, and electric vehicle more data than a home does, and remote patients. Secure, low-latency
charging stations, etc.
32 5G Edge Automation & Intelligence| November 2021

connectivity, wearable devices, smart automation environments. new experiences to the users. Edge
in-home sensors, and computer vision automation can therefore lower the
make virtual health management 3.6 Cloud and Edge Gaming latency, as well as manage additional
possible while reducing costs and latency requirements added by
increasing accessibility to healthcare. The demand for high quality, additional network or compute cycles.
Additionally, experiments with remote high throughput, and low latency It provides a path to bring a true end-
surgery show promising results while is continuously increasing as a to-end gaming experience that can
edge-based computer vision enable requirement for gaming platforms. leverage edge to perform a hybrid
flagging of health conditions in real- For example, modern games expect of frame streaming and command
time. Edge automation enables these gaming infrastructure to support vast streaming models to deliver highly
types of solutions to make better amount of data processing to render responsive gaming experience without
and faster decisions while keeping frames at the highest quality at the sacrificing quality and frame rate.
computer vision models local where highest frame rate possible, support
possible therefore maintaining for hardware-accelerated ray tracing,
and AI capabilities plus a vast amount 3.7 Scalable Digital Twin
compliance with privacy regulations.
of storage space (e.g., 150GB of A Digital Twin (DT) is a real-time virtual
storage space per game install). representation of a physical entity
3.5 Enabling Location such as an object, a system, or a
Information The introduction of cloud gaming process. Using connected sensors,
platforms enabled gamers to play on
Location information is essential for this cyber-physical technology permits
any device connected to the Internet
key services and applications with connectivity and synchronization
without the need to upgrade their
use cases such as precise e911, between the physical components
client gaming platforms. Cloud gaming
fraud prevention and mitigation, and their digital counterparts. Further,
offers two models of operation,
hyper-local customer applications through analytics and simulations
frame streaming and command
such as weather, instant couponing or using the digital model, the Digital
streaming model. The frame streaming
experience zones. It is also used for Twin technology can produce future
model eliminates the need for high-
real-time network optimization. predictions with rich insights about the
performance GPU and storage in the physical entity.
While Global Positioning System client platform by rendering the game
(GPS) or device-provided location is in the cloud instance and stream The unique characteristics of DT
commonly used, uplink-based location the encoded frames to the client technology has several potential
tracking has the advantage of better device. The command streaming applications in the fields of
UE battery management and user model eliminates the storage needs infrastructure, smart cities,
control of the location info. at the client by launching the game manufacturing, natural resources,
in the cloud instance and stream the healthcare, etc. For example, in
For the uplink-based method, Next GPU APIs and data to render a frame ITS, DTs can accurately simulate
Generation (NG)-RAN, an O-RAN using the client GPU. Though the transport network of a city and can
node includes Next Generation frame streaming model offers better optimize traffic efficiency, planning
NodeB (gNB) and real- and non-real- throughput, the end-to-end user input and development of transport
time RICs that enhance traditional to display latency in cloud gaming is infrastructure. In manufacturing, the
network functions with embedded much higher than the client gaming DT technology has demonstrated
intelligence. The gNB controls multiple platforms and potentially degrades disruptive impact on handling complex
Transmission Reception Points the overall gaming experience. processes like product lifecycle
(TRPs) an antenna array with two or Additionally, the transcoding and management, asset maintenance,
more antenna elements located at streaming content from the cloud production line efficiency optimization,
a specific geographical location for to the client under varying network etc.
an area. Using Sounding Reference conditions brings visual quality
Signal (SRS) measurements at So far, the DT technology has seen
variations.
different TRPs, the uplink location can limited adoption in the industry
be calculated in real-time at a high As low latency requirements, link due to its stringent requirements
level of accuracy – as low as sub-20 variability, and available bandwidth on communication and computing
meters. are still challenging, new approaches infrastructure. In large scale
related to the 5G edge and MEC space applications such as in smart city, ITS,
In addition to the mainstream use are being considered to dynamically or manufacturing, the DT technology
cases, this method enables velocity distribute rendering between cloud, needs to collect data from a large
measurement that could provide edge, and client based on latency number of connected sensors, and it
crash detection capabilities as well budget, bandwidth, or acceleration is not yet viable to transport all these
as angle of arrival as an alternate requirements of the gaming workload. data to the data center for processing.
method to barometric pressure for The 5G plane enables a new way of On the other hand, DT also requires
z-axis positioning which are critical for accessing streamed game content powerful computation resources to
applications in automotive or industrial inside and outside the home giving analyze the sensor data, simulate
33 5G Edge Automation & Intelligence| November 2021

complex digital models, and generate future predictions in real-time.

The combination of 5G and edge technologies can provide feasible and cost-
effective solutions to realize DT applications. Edge computing allows processing
of sensor data close to the source, thus avoiding the need to transport large
amounts of data over long distances. Alongside, the low-latency communications
offered by 5G can help to achieve real-time service requirements of the DT
applications. On the computational front, microservices based architecture can
provide a scalable and flexible solution for the DT system to keep up with the
analysis and simulation tasks in real time.
34 5G Edge Automation & Intelligence| November 2021

Conclusion

5G and edge computing are two intertwined technologies that will converge and work together to significantly improve
the performance of applications and enable massive amounts of data to be processed in near real-time at different
locations (edge zones). Edge automation and optimization with AI/ML can help to automate and optimize network system
processes and service delivery at each available edge zone or throughout multiple edge zones. The ultimate goals of this
symbiosis between 5G and edge involve increased performance guarantees, enhanced workload balancing, improved
processing capabilities and times via 5G edge automation and optimization, reduced human intervention up to zero-touch
management and orchestration. Edge intelligence, based on the low latency, high reliability 5G connection at edge and the
AI/ML processing power provided by edge computing enables pervasive intelligence on all connected edge devices, as well
as distributed data analysis and distributed learning on connected edge devices.
Equipped with AI/ML-driven capabilities, the 5G edge can be further augmented. Differentiation is crucial between AI/
ML-based solutions for the network to control, manage, and orchestrate resources and functions and how systems should
be designed to improve the performance and resource utilization of AI/ML-based solutions. Integrating AI/ML advances to
5G edge automation will enrich human experience, enable autonomic decision-making with adaptive policies, and reduce
or eliminate human errors. The implementation of AI/ML-driven optimization at the 5G edge enables the adaptation to
evolving situations, conditions, and diverse environments as well as seeing through the complex interactions between
various components of a system and optimizing the resource usage.
5G edge automation and optimization can enhance the 5G edge with various new features and enable multiple key
technologies. In this white paper, some key features and technologies related to data collection and processing, context
discovery and situational awareness, how to handle (network) dynamics, explainable AI, multi-access, distributed learning,
and achieving the joint optimization of communication and computing have been discussed. An analysis of requirements
followed discussion around potential directions for network architecture to demonstrate the gaps and needs for enabling
the discussed key features and technologies. Finally, a list of selected use cases demonstrates key benefits and
challenges facing industries today regarding 5G edge automation and optimization.
35 5G Edge Automation & Intelligence| November 2021

Acronyms

3GPP: 3rd Generation Partnership EN-DC: Evolved-Universal Terrestrial ITS: Intelligent Transportation

Project  Radio Access New Radio Dual Systems 
Connectivity 
5G: 5th Generation  ETSI: European Telecommunications LCM: Life Cycle Management 
Standards Institute 
AI: Artificial Intelligence  FPGA: Field Programmable Gate LF: Linux Foundation
Arrays 
IAIOps: AI Operations  LFU: Least-Frequently-Used 
GMA: General Multi-Access 
AIMD: Additive Increase/
LLS: Lower-Layer Split 
Multiplicative Decrease  gNB: Next Generation NodeB 

API: Application Program Interface  LRU: Least-Recently-Used 

GPU: Graphical Processing Unit 
MANO: Management and
AR: Augmented Reality 
GPS: Global Positioning System  Orchestration 

BBU: BaseBand Unit  GSMA: Global System for Mobile MDA: Management Data Analytics 
Communications Alliance 
C-SON: Centralized SON  MEC: Multi-Access Edge Computing 
HCP: Hyperscale Cloud Provider 

CEI: Collaborative Edge Intelligence  HVAC: Heating, Ventilation, and Air MEF: Mobile Ecosystem Forum 
Conditioning 
CNF: Cloud-native Network Function  ML: Machine Learning 
I2P: Infrastructure-to-Pedestrian 

CPU: Central Processing Unit  MNO: Mobile Network Operator 

IaaS: Infrastructure-as-a-Service 

COTS: Commercial Off-The-Shelf  MR: Machine Reasoning 

IAB: Integrated Access and Backhaul 

CSP: Cloud Service Provider  NAT: Network Address Translation 

ICN: Information Centric Networking 
CSMF: Communication Service
NDN: Named Data Networking 
Management Function  ICS: Industrial Control Systems 

CU: Centralized Unit  IEEE: Institute of Electrical & NFV: Network Function Virtualization 
Electronic Engineers 
DNN: Deep Neural Network  IID: Independent and Identically NG: Next Generation 
Distributed 
D-SON: Distributed SON  NR: New Radio 
IoT: Internet of Things 
DSRC: Dedicated Short-Range
NSI: Network Slice Instance 
Communications  IP: Internet Protocol 
NSMF: Network Slice Management
DT: Digital Twin 
IPSec: Internet Protocol Security  Function 
NSSMF: Network Slice Subnet
DU: Distributed Unit  ISR: Intelligence, Surveillance, and Management Function 
Reconnaissance 
NWDAF: Network Data Analytics
ECaaS: Edge Compute-as-a-Service 
IT: Internet Technology  Function 
36 5G Edge Automation & Intelligence| November 2021

Acronyms

OAM: Operations Administrations

TIP: Telecom Infrastructure Project 
and Management 
O-Cloud: Orchestrator and Cloud TLRU:
Platform  Time-aware Least-Recently-Used 
ONAP: Open Networking Automation
TM: Traffic Management 
Platform 

ONF: Open Networking Foundation  TMF: Tele-Management Forum 

O-RAN: Open Radio Access Network  TRP: Transmission Reception Points 

OT: Operational Technology  TSG: Technical Specification Group 

PaaS: Platform-as-a-Service  TSN: Time Sensitive Networks 

PCF: Policy Control Function  UDP: User Datagram Protocol 

PoP: Point of Presence  UE: User Equipment 

QoS: Quality of Service  UP: User Plane 

RAN: Radio Access Network  UPF: User Plane Function 

RAT: Radio Access Technology  UWB: Ultra-Wide Band 

RIC: RAN Intelligent Controller  V2I: Vehicle-to -Infrastructure 

RSU: RoadSide Units  VNF: Virtual Network Function 

RU: Radio Unit  VPN: Virtual Private Network 

SDN: Software Defined Network  VR: Virtual Reality 

SFG: Security Focus Group  VRU: Vulnerable Road User  

SI: System Integrator  XAI: Explainable AI 

ZSM: Zero-Touch Network and

SLA: Service Level Agreement 
Service Management 
SON: Self-Optimizing/Organizing
Network 

SRS: Sounding Reference Signal 

TCP: Transmission Control Protocol 

37 5G Edge Automation & Intelligence| November 2021

References

[1] 3GPP, System Architecture for the 5G System (5GS), V.17.1.1, 3GPP TS 23.501, June 2021.
[2] 3GPP, Study on management aspects of edge computing, V16.0.1, 3GPP TR 28.803, Sept. 2019.
[3] 3GPP, 5G System Enhancements for Edge Computing, V1.0.0,, 3GPP TS 23.54, June 2021.
[4] 3GPP, Study on enhancements of edge computing management, V1.0.0, 3GPP TS 28.814, June 2021.
[5] 3GPP, Architecture for enabling Edge Applications, V.17.0.0, 3GPP TS 23.558, June 2021.
[6] 3GPP, Architecture enhancements for 5G System (5GS) to support network data analytics services, V17.1.0,
3GPP TS 23.288, June 2021.
[7] 3GPP, Study on enablers for network automation for the 5G System (5GS)”, V.17.0.0, 3GPP TR 23.700-91, Dec.
2020.
[8] 3GPP, Study on enhancement of Management Data Analytics (MDA), V17.0.0, 3GPP TR 28.809, March 2021.
[9] 3GPP, Management and orchestration; Management Data Analytics (MDA), V0.0.0, 3GPP TS 28.104, April 2021.
[10] 3GPP, Study on enhancement for Data Collection for NR and EN-DC, V.0.1.0, 3GPP TR 37.817, Jan. 2021.
[11] 3GPP, Management and orchestration; Edge Computing Management,, 3GPP TS 28.538, June 2021.
[12] 3GPP, Study on Security Aspects of Enhancement of Support for Edge Computing in 5GC, V0.6.0, 3GPP TR
33.839, May 2021.
[13] 3GPP, 5G System; Network Data Analytics Services; Stage 3, V.17.3.0, 3GPP TS 29.520, Jun. 2021.
[14] O-RAN, “O-RAN Architecture Description”, v05.00, O-RAN, WG1, 2021.
[15] T. Forum. [Online]. Available: https://www.tmforum.org/catalysts/the-edge-in-automation/.
[16] T. Forum. [Online]. Available: https://www.tmforum.org/collaboration/catalyst-program/
artificial-intelligence-operations-aiops/.
[17] ETSI, “Zero-touch network and Service Management (ZSM); Reference Architecture, ETSI GS ZSM 002 V1.1.1,”
2019.
[18] ETSI, “Zero-touch network and Service Management (ZSM); Closed-Loop Automation; Part 1: Enablers, ESTI
ZSM009-1 V1.1.1,” 2021.
[19] Apache, “Apache Kafka,” [Online]. Available: https://kafka.apache.org/.
[20] Apache, “Apache Pulsar,” [Online]. Available: https://pulsar.apache.org/.
[21] RabbitMQ, “Rabbit-MQ,” [Online]. Available: https://www.rabbitmq.com/.
[22] Apache, “Apache Spark,” [Online]. Available: https://spark.apache.org/.
[23] 3GPP, “”Study on management and orchestration of network slicing for next generation network”, V15.1.0, 3GPP
TR 28.801,” 2018.
[24] Omdia, “Artificial Intelligence for Edge Devices,” Available: https://omdia.tech.informa.com/OM011942/Artificial-
Intelligence-for-Edge-Devices, 2020.
[25] P. S. Dutta, N. R. Jennings and L. Moreau, “Cooperative Information Sharing to Improve Distributed Learning in
Multi-Agent Systems,” Journal of Artificial Intelligence Research, vol. 24, p. 407–463, 2005.
[26] V. I. Bajić, L. Weisi and T. Yonghong, “Collaborative intelligence: Challenges and opportunities,” in IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021.
[27] Ericsson, “Edge Computing and 5G,” [Online]. Available: https://www.ericsson.com/49d80b/assets/local/digital-
services/trending/edge-computing/edge-computing-5g-report.pdf.
[28] J. Zhang and K. B. Letaief, “Mobile Edge Intelligence and Computing for the Internet of Vehicles,” Proceedings of
the IEEE, vol. 108, no. 2, pp. 246-261, 2020.
[29] 3GPP, “”Overall Description of Radio Access Network (RAN) Aspects for Vehicle-to-Everything (V2X) based on LTE
38 5G Edge Automation & Intelligence| November 2021

and NR”, 3GPP TR 37.985, V.16.0.0,” 2020.

[30] 3GPP, “”Integrated Access and Backhaul Radio Transmission and Reception”, 3GPP TS 38.174, V.16.3.0,” 2021.
[31] 3GPP, “”Study on RAN-centric data collection and utilization for LTE and NR”, 3GPP TR 37.816, V.16.0.0,” 2019.
[32] ETSI, “Mobile Edge Computing (MEC); Framework and Reference Architecture, ETSI GS MEC 003, V2.1.1,” Jan.
2019.
[33] ETSI, “MEC in 5G Networks, ESTI White Ppaer No.28,” Jun. 2018.
[34] 3GPP, “3rd Generation Partnership Project; Technical Specification Group Services and System Aspects; System
Architecture for the 5G System; Stage 2 (Release 15),” June 2018.
[35] Ericsson, “Explainable AI – How Humans Can Trust AI,” [Online]. Available: https://www.ericsson.com/en/
reports-and-papers/white-papers/explainable-ai--how-humans-can-trust-ai.
[36] ETSI, “MEC Framework and Reference Architecture,” [Online]. Available: https://www.etsi.org/deliver/etsi_gs/
MEC/001_099/003/02.01.01_60/gs_MEC003v020101p.pdf.
[37] ETSI, “Traffic Management APIs,” [Online]. Available: https://www.etsi.org/deliver/etsi_gs/
MEC/001_099/015/02.01.01_60/gs_mec015v020101p.pdf.
[38] R. 8743, “Multi-Access Management Service,” [Online]. Available: https://www.rfc-editor.org/rfc/rfc8743.txt.
[39] 3GPP, “5G System; Access Traffic Steering, Switching and Splitting (ATSSS); Stage 3,” [Online]. Available: https://
portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3607.
[40] IETF, “RFC 2890 “Key and Sequence Number Extensions to GRE”,” [Online]. Available: https://www.rfc-editor.org/
rfc/rfc2890.txt.
[41] K. Sundaresan, S. Park and R. Sivakumar, “Transport Layer Protocols in Ad Hoc Networks,” in Ad Hoc Networks,
Springer, Boston, MA, 2005.
[42] I. K. a. K. Wehrle, “Transport Protocol Issues of In-Network Computing Systems,” IETF, 2020.
[43] L. Peterson, T. Anderson, S. Katti, N. McKeown, G. Parulkar and J. Rexford, “Democratizing the Network Edge,” in
SIGCOMM CCR, 2019.
[44] Y. Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars and L. Tang, “Neurosurgeon: Collaborative Intelligence
Between the Cloud and Mobile Edge,” in 22nd International Conference on Architectural Support for Programming
Languages and Operating Systems, 2017.
[45] M. Eisen, K. Arjun, A. S. Baxi and D. Cavalcanti, “Network Performance Adaptation in Wireless Control with
Reinforcement Learning,” in 54th Asilomar Conference on Signals, Systems, and Computers, 2020.
[46] P. Kairouz, H. B. McMahan and e. al., “Advances and open problems in Federated Learning,” Foundations and
Trends in Machine Learning, vol. 14, no. 1-2, 2021.
[47] R. Balakrishnan, M. Akdeniz, S. Dhakal, A. Anand, A. Zeira and N. Himayat, “Resource Management and Model
Personalization for Federated Learning over Wireless Edge Networks,” Journal of Sensor and Actuator Networks,
vol. 10, no. 17, 2021.
[48] A. Anand, S. Dhakal, M. Akdeniz, B. Edwards and N. Himayat, “Differentially Private Coded Federated Linear
Regression,” in IIEE SP Data Sciences and Learning workshop, 2021.
[49] M. Isaksson and K. Norrman, “Secure Federated Learning in 5G Mobile Networks,” in IEEE Global
Communications Conference, 2020.
[50] S. Niknam, H. S. Dhillon and J. H. Reed, “Federated Learning for Wireless Communications: Motivation,
Opportunities, and Challenges,” IEEE Communications Magazine, vol. 58, no. 6, pp. 46-51, 2020.
[51] Ericsson, “Edge computing and deployment strategies for communication service
providers,” [Online]. Available: https://www.ericsson.com/en/reports-and-papers/white-papers/
edge-computing-and-deployment-strategies-for-communication-service-providers.
[52] Intel, “Intel’s Converged Edge Insights,” [Online]. Available: https://www.intel.com/content/www/us/en/edge-
computing/edge-software-hub-converged-edge-insights.html.
[53] J. Research, “INDUSTRIAL IOT: FUTURE MARKET OUTLOOK, TECHNOLOGY ANALYSIS & KEY PLAYERS 2020-2025,”
39 5G Edge Automation & Intelligence| November 2021

Oct. 2020.
[54] OPC, “OPC UA,” [Online]. Available: https://opcfoundation.org/about/opc-technologies/opc-ua/.
[55] IETF, “ICN IETF,” [Online]. Available: https://datatracker.ietf.org/rg/icnrg/about/ .
[56] ISA, “ISA 95,” [Online]. Available: https://www.isa.org/isa95/.
[57] C. Cities, “https://www.c40.org/why_cities”.
[58] WHO, “Road traffic injuries, https://www.who.int/health-topics/road-safety#tab=tab_1,” WHO, 2021.
[59] Intel, “The Future of Smart Road Infrastructure,” [Online]. Available: https://www.intel.com/content/www/us/en/
transportation/smart-road-infrastructure.html.
[60] J. Clayton, “Crafting a Powerful Executive Summary,” Harvard Business School, 8 Sept. 2003. [Online]. Available:
https://hbswk.hbs.edu/archive/crafting-a-powerful-executive-summary. [Accessed 13 Feb. 2020].
[61] O. Mayr, The Origins of Feedback Control, Clinton, MA USA: The Colonial Press, Inc., 1970.
[62] L. Zhang, A. Afanasyev, J. Burke, V. Jacobson, K. Claffy, P. Crowley, C. Papadopoulos, L. Wang and B. Zhang,
“Named Data Networking,” in ACM SIGCOMM Computer Communication Review (CCR), July 2014.
[63] C. Yi, J. Abraham, A. Afanasyev, L. Wang, B. Zhang and L. Zhang, “On the Role of Routing in Named Data
Networking,” in ACM Conference on Information-Centric Networking, 2014.
[64] R. Pirmagomedov, S. Srikanteswara, D. Moltchanov, G. Arrobo, Y. Zhang, N. Himayat and Y. Koucheryavy,
“Augmented Computing at the Edge Using Named Data Networking,” in IEEE Globecom Workshops (GC Wkshps),
Dec. 2020.
[65] L. Zhang, A. Afanasyev, J. Burke, V. Jacobson, K. Claffy, P. Crowley, C. Papadopoulos, L. Wang and B. Zhang,
“Named Data Networking,” in ACM SIGCOMM Computer Communication Review (CCR), 2014.
[66] M. Polese, R. Jana, V. Kounev, K. Zhang, S. Deb and M. Zorzi, “Machine Learning at the Edge: A Data-Driven
Architecture with Applications to 5G Cellular Networks,” IEEE Transactions on Mobile Computing, vol. doi:
10.1109/TMC.2020.2999852.
[67] S. Prakash, S. Dhakal, M. Akdeniz, Y. Yona, S. Talwar and N. Himayat, “Coded computing for low latency Federated
Learning over Wireless Edge Networks,” IEEE Journal of Selected Areas in Communication, Special issue on ML
for Communications and Networking, vol. 39, no. 1, 2021.
40 5G Edge Automation & Intelligence| November 2021

Acknowledgments
5G Americas’ Mission Statement: 5G Americas facilitates and advocates for
the advancement and transformation of LTE, 5G and beyond throughout the
Americas.
5G Americas’ Board of Governors members include Antel, Airspan Networks,
AT&T, Cable & Wireless, Ciena, Cisco, Crown Castle, Ericsson, Intel, Liberty
Latin America, Mavenir, Nokia, Qualcomm Incorporated, Samsung, Shaw
Communications Inc., T-Mobile USA, Inc., Telefónica, VMware and WOM.
5G Americas would like to recognize the significant project leadership and
important contributions of group leaders Meryem Simsek of VMWare and Clark
Chen of Intel along with many representatives from member companies on 5G
Americas’ Board of Governors who participated in the development of this white
paper.
The contents of this document reflect the research, analysis, and conclusions
of 5G Americas and may not necessarily represent the comprehensive opinions
and individual viewpoints of each particular 5G Americas member company.
5G Americas provides this document and the information contained herein for
informational purposes only, for use at your sole risk. 5G Americas assumes
no responsibility for errors or omissions in this document. This document is
subject to revision or removal at any time without notice. No representations
or warranties (whether expressed or implied) are made by 5G Americas and
5G Americas is not liable for and hereby disclaims any direct, indirect, punitive,
special, incidental, consequential, or exemplary damages arising out of or in
connection with the use of this document and any information contained in this
document.
© Copyright 2021 5G Americas