Page 125 - 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

In the field of cross-slice optimization, different Another important novelty comes from the
approaches exist exploiting several ML tools. specification of the SLA terms for a RAN slice to be
Q-learning was used in [23] to design a slicing used by the ML-based solution. This paper takes as
controller that decides which resource units are a reference the attributes defined in the GSMA
allocated to each slice based on requirements at the Generic Slice Template considered by 3GPP to
user level. Q-learning complemented with a genetic specify the SLA to be fulfilled for a RAN slice across
algorithm was considered in [24] for scaling down a geographical area covering multiple cells in terms
allocated resources to slices for congestion control of, e.g. the total amount of capacity to be provided
purposes. In [25] deep deterministic policy gradient to each slice. Instead, other approaches such as
(DDPG) is used to allocate resource blocks to [28]-[32] just consider the SLA specified in terms of
different tenants in a cloud RAN environment. In the QoS parameters defined at the user level, but
turn, game theory with exponential learning is without enforcing any aggregate capacity per slice.
proposed in [26] to divide the network resources Finally, another difference with respect to previous
(i.e. bandwidth) among slices using OpenFlow, works comes from the algorithmic solution
being a general approach not particularized to the considered in the proposed framework, which
specificities of radio resource allocation. Recently, consists of a multi-agent DQN with one agent per
deep Q learning has become a quite popular tool for slice that learns the capacity to be allocated to each
allocating radio resources to slices, as reflected by slice in each cell. In contrast to single agent
works [27]-[33] that include different variants of solutions like those of [30], [31], which jointly
this technique and address the problem from consider all the tenants when making decisions, the
different perspectives, such as the joint allocation of multi-agent approach has advantages such as better
computational resources and radio resources to scalability as it allows easily adding/removing
users in [27], the allocation of aggregate capacity slices in the scenario simply by adding/removing
per slice to multiple cells in [28], [29], the allocation the corresponding agent. Moreover, while some
of resources to slices on a single cell basis in [30], multi-agent approaches have already been
[31], [32], or the allocation of per-cell resources to considered in [28], [29][32], the one considered
the different slices jointly considering multiple cells here has the advantage that an agent learns the
in [33]. Finally, other works have proposed the use policy for assigning capacity to be provided to the
of traffic forecasting for cross-slice resource slice in each cell, in contrast to [32], which
allocation, applying techniques such as LSTM neural considered the capacity in a single cell, or [28], [29],
networks [34], deep convolutional neural networks which provided the aggregated capacity over all the
[35], Generative Adversarial Networks (GANs) [36], cells.
or deep neural networks [37].
3. ML-ENABLED CROSS-SLICE
This paper introduces several novelties with
respect to previous works. First of all, this paper MANAGEMENT FRAMEWORK
presents a functional framework aligned with 3.1 O-RAN framework for ML-assisted
current 3GPP and O-RAN specifications for solutions
implementing ML-assisted cross-slice radio
resource optimization and particularizes it to a As part of the specification of new interfaces and
specific algorithmic solution coming from our functionality for an open and intelligent RAN, the O-
previous work [33]. Instead, the above-mentioned RAN Alliance is working on the definition of a
works have put the focus on algorithm development framework for the deployment of ML-assisted
but without going into detail of the mapping on solutions within the RAN (i.e. solutions that rely on
existing architectures from standardization bodies. the use of ML models such as supervised learning,
For example, some works just consider a slicing reinforcement learning, etc.) [38].
controller (e.g. [23]) or a network slicing module A representation of the overall RAN functional
(e.g. [28], [29]) but without providing details of how architecture being defined by O-RAN is illustrated
this would be mapped on practical architectures. in Fig. 1 [39]. This constitutes a disaggregated RAN,
Only in [24] an architectural framework for slice compliant with 3GPP specifications, where the
management and orchestration that is aligned with radio protocol stack is split and distributed between
3GPP is presented, but without providing specific different RAN nodes. In particular, the O-RAN Radio
details on the algorithm implementation. Unit (O-RU) hosts the RF processing and the lower
part of the PHY layer functionality (e.g. i/FFT

© International Telecommunication Union, 2020 105