Page 127 - ITU Journal, Future and evolving technologies - Volume 1 (2020), Issue 1, Inaugural issue

ITU Journal on Future and Evolving Technologies, Volume 1 (2020), Issue 1

runtime environment where offline training of the into other network components via the different
ML model takes places. This refers to the training of specified interfaces. For example, management
model before being executed within the network. In configuration actions from an actor within the SMO
addition to the data collected from the real network, layer on any RAN node can be conducted via the O1
offline training may also rely on synthesized data interface, control actions on O-CU/O-RU from an
which can accurately reproduce the behavior of the actor within the near-RT RIC can go over the E2
real network environment. The training may interface and policy management configuration
include an evaluation stage to assess the actions between the non-RT RIC and the near-RT
performance of the model and validate that it is RIC can be communicated over the A1 interface.
ready and reliable to be deployed in the live
network environment. Offline training is necessary 3.2 Information models for network slice
to obtain supervised learning models (e.g. deep management
neural networks, support vector machines, etc.) as With regard to the management of network slicing
well as reinforcement learning models (e.g. Q- in 5G networks, 3GPP specifications include
learning, multi-armed bandit learning, deep RL). information model definitions, referred to as
The training host component is likely to be part of Network Resource Models (NRMs), for the
the SMO layer. characterization of network slices [42] together
The ML inference host represents the runtime with a set of management services (MnS) for
environment where the (previously trained and network slice life-cycle management (e.g. network
validated) ML model is executed and fed with online slice provisioning MnS for network slice creation,
data to produce the outputs that will be used in the modification and termination, performance
network operation. Multiple ML inference hosts can monitoring services per slice, etc.) [43]. In addition,
be in place, whose location depends on aspects such work is being conducted at 3GPP level to support
as the purpose and type of ML models being SLA/SLS management [44], as well as closed loop
executed, its computation complexity, the assurance solutions that allow a service provider to
availability and the quantity of data used and the continuously deliver the expected level of
response time requirements (real-time or non-real- communication service quality in a 5G network [45].
time) of the ML application. Hence, ML inference Fig. 3 provides an overview of the different types of
hosts can be placed within the SMO layer but also information models and their relations that are
within the RAN nodes (i.e. near-RT RIC, O-CU, O-DU). relevant for network slice management. The main

In turn, the actor represents the network entity (i.e. idea behind the overall flow of the information
UE, O-DU, O-DU, Near-RT RIC and Non-RT RIC) that models, as illustrated in Fig. 3, is that a network slice
hosts the decision-making function that consumes is conceived as a “product” offered by a Network
the outputs of the ML inference host and takes Slice Provider (NSP) to a Network Slice Customer
actions. It is worth noting that the distinction (NSC). In this respect, the GSMA Generic Slice
between the ML inference host and the actor obeys Template (GST) is used as the SLA information
the fact that these components may or may not be associated with the network slice product for the
co-located as part of the same network entity. An communication between the NSC and NSP through,
example of non-co-location could be the case of a e.g. a Business Support Systems (BSS) product
mobility prediction model executed in an inference order management Application Programming
host within the non-RT RIC that produces outputs Interface (API).
(e.g. mobility patterns) that are retrieved and The GSMA GST provides a standardized list of
consumed by the near-RT RIC (i.e. the actor in this attributes (e.g. performance related, function
case) for enhanced RRM (e.g. handover decisions related, etc.) that can be used to characterize
based on mobility patterns). In contrast, an example different types of network slices [46]. GST is generic
of co-location could be an RRM algorithm for and is not tied to any type of network slice or to any
mobility management that embeds a reinforcement agreement between an NSC and an NSP. A Network
learning model and is executed within the near-RT Slice Type (NEST) is a GST filled with (ranges of)
RIC, which in this case serves as both the inference values. There are two kinds of NESTs: Standardized
host and the actor. The actions decided by the actor NESTs (S-NEST), i.e. NESTs with values established
can be handled either internally within the actor by standards organizations, working groups, fora,
(e.g. RL-based RRM algorithm for mobility etc. such as, e.g. 3GPP, GSMA, 5GAA, 5G-ACIA, etc.;
management within the near-RT RIC) or enforced and Private NESTs (P-NEST), i.e. NESTs with values

© International Telecommunication Union, 2020 107