Page 119 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

transmission technologies (ZigBee, Bluetooth, and Low- of the spectrum entails a cost that may turn unaffordable as
Power WiFi), LPWAN, and cellular networks from 2G to 5G the number of deployed IoT devices grows; either due to the
are considered. The wide diversity of RATs available on the expenses derived from renting the infrastructure to the MNO
market is remarked, as well as the necessity for finding a way or due to the direct cost of licensing the bands from the
to combine them. competent authorities. Even if data transmissions are
sporadic, the total aggregated traffic of the network may
On the other hand, preliminary works about multi-RAT render cellular-based IoT deployments hard to maintain in
support for 5G can be found in the literature. Authors in [14] terms of operational costs. On the other hand, and although
bring to light the need of heterogeneous Radio Access 5G is envisioned to reduce power consumption, cellular-
Networks (RAN) to alleviate the congestion and overload of based nodes are intrinsically dependent on signaling. This
the cellular infrastructure. The mixture of cellular networks overhead in the communications unavoidably lead to an
with RATs working on ISM frequency bands, such as increase in power consumption of IoT devices [18].
WLAN technologies, is proposed for the forthcoming 5G.
This concept is also remarked in [15], where authors claim Under these circumstances, the use of multiple RATs is one
that the intelligent integration of WiFi and cellular networks of the key tools of 5G deployments to benefit from the
can duplicate or even triplicate the quality of service and advantages of each wireless technology. However, the
network performance. scientific community has focused on a set of supporting
technologies (WiFi/WiMAX/mm-wave/etc.) potentially ill-
In this framework, very few papers apply ML techniques to suited to IoT devices [19], [20]. First, due to the large power
smartly optimize the access to the medium in 4G/5G for consumption of such technologies (in particular when
M2M communication. Authors in [16] proposed an RL- compared to LPWAN alternatives), and second due to the
based algorithm for cellular networks that enables Machine cost increase of their respective radio transceiver (especially
Type Communication devices to cooperatively communicate for WiMAX/mm-wave and, again, when compared to
to minimize the network congestion. This is accomplished LPWAN technologies such as LoRa). Furthermore, previous
by the intelligent selection of the base station to transmit studies in this field have solely focused on alleviating
from the device-side. The use of ML techniques is also network congestion, neglecting the intrinsic requirements of
exploited in [17], where authors introduced an ant-colony IoT deployments, that is, their low-cost and low-power
heuristic algorithm to smartly decide which RAT should be nature.
used by users. These decisions were made to maximize
system utility and better balance resource utilization. The As 5G ultimately aims to encompass a wide variety of traffic-
RATs considered were LTE, WiMAX, and WiFi and the generating devices, the authors believe that the future 5G
results obtained showed a performance improvement standard releases should devote further efforts to
between 20% and 70% with respect to other RAT usage acknowledge the limitations of IoT networks. In this paper
strategies. However, none of these ML works consider the we have made a step forward in this direction, by not only
nature of IoT devices (low-power consumption and limited maximizing throughput of IoT nodes, but also considering
hardware resources) or the use of LPWAN technologies potential restrictions in the usage of the different RATs (such
(which are known to be well-suited for IoT devices). as battery limitation, daily transference quota, etc.) In this
sense, we believe, and have demonstrated in this work, that
3. IMPORTANCE OF THE PROPOSED SOLUTION Machine Learning techniques can play an essential role in
TO THE 5G STANDARDIZATION PROCESS deriving optimal transmission policies for the future 5G.
Therefore, we envisage that 5G could potentially benefit
As shown in previous sections, the use of cellular networks from this subfield of the Artificial Intelligence area and
is gathering momentum as an enabler for the IoT due to the hence, should be paid more attention by the standardization
need for global coverage in most user and industrial bodies.
applications. This requirement can be met by employing the
infrastructure of the Mobile Network Operators (MNO), 4. MATHEMATICAL FRAMEWORK
which already offers an almost-global coverage. This
solution also reduces the installation cost of IoT systems by As discussed above, the objective of the policy-derivation
avoiding the acquisition of specific equipment to connect algorithm is to determine which RAT should be used by an
IoT devices to the Internet. In this context, several efforts are IoT device in any given situation. This can be translated into
devoted to make cellular networks more suitable for the IoT, the RL jargon as determining the optimal action to take
leading to the emergence of the 5th Generation of mobile (out of a set of allowed actions, being ∈ ) given a state
networks (5G). 5G is envisaged to adapt the advantages of (a description of the internal/external state of the IoT node,
cellular networks to the characteristics of the IoT, that is, with ∈ ). Having performed action in the state , some
massive number of devices, enabling lower end-to-end reward is obtained -this feedback signal helps nodes
latency and energy consumption, and global coverage [7]. understand what actions are better to take than others-. This
reward can, for instance, measure how much information has
However, there are still some shortcomings that may delay been reported, how important such information was, etc. The
the ubiquitous use of 5G in IoT. The first stems from the cost function ℛ mathematically defines such a reward as a
of using licensed frequency bands. Using such private parts

– 103 –