Page 118 - Proceedings of the 2018 ITU Kaleidoscope

2018 ITU Kaleidoscope Academic Conference

Conversely, cellular-oriented communication protocols, like sporadic traffic and very low energy consumption (cellular
LTE and the much-anticipated 5G, are geared towards faster, technologies) or to a low offered bitrate (LPWAN).
more stable transmissions of information that, indeed, can be Therefore, there is an increasing trend toward the use of
of interest to certain situations -e.g. to transmit a daily report multiple RATs in the IoT to minimize this problem.
of an important asset, or in general, to deliver large packets-. However, although many works look into employing
For all other situations, and with the aim of maximizing heterogeneous RATs in wireless networks (as will be
network and cost efficiency, a promising suite of solutions commented in the Related Work section), to the best of our
that make use of the unlicensed ISM (Industrial, Scientific knowledge, very few studies tackled this problem from a
and Medical) bands is being extensively used as a RAT for mathematical point of view –that is, how nodes optimally
the IoT. As one of the main key features of Smart decide which RAT should be used at any given time–. With
Cities/Industries is the large areas to be covered, low-power the interest of filling this gap, and by making use of the
wide area networks (LPWAN) technologies are called to be mathematical framework of Reinforcement Learning (RL), a
the de-facto mean of communication for those small and subfield of Machine Learning (ML), we propose a
sporadic transferences of information. The distinguishing mechanism to derive transmission policies that optimally
feature of LPWAN is the long-range radio links forming a determine the RAT to be employed for each transmission.
star network topology, in which end devices (nodes) are These transmission policies consider the global state of the
directly connected to a collector device (gateway) that node and are oriented to maximize some predefined
provides access to the IP network. These networks are performance metric while being computable in very
designed to notably improve the battery life of nodes and hardware-constrained devices.
support on-demand bursty traffic by reducing the signaling
overhead to a minimum. Therefore, the contribution of this work is threefold: (i) a
methodological and thorough justification for the need of
However, the main drawback of LPWAN technologies is its multiple RATs in IoT-oriented 5G networks; (ii) a
low bitrate. Depending on the technology and configuration, mathematical formulation that models performance of IoT
it fluctuates from one hundred bits per second to a few nodes as an RL problem that truly embraces the nature of
thousands of bits per second, potentially being insufficient these devices; and (iii) the proposal of a state-of-the-art ML
for handling all the traffic generated by an IoT device. Such technique to solve such an RL problem and its analysis, via
low bitrate is due to the modulation techniques employed in simulation, to highlight the importance of ML-oriented
most LPWAN, that focus on being as much robust as transmission policies in 5G deployments.
possible against interferences and increasing the receiver
sensitivity (depending on the specific technology, up to -150 The rest of the paper is organized as follows. In Section 2, a
dBm) to achieve large link distances. Therefore, a trade-off review of the related work is presented. Next, in Section 3
between data rate and sensitivity arises. For instance, in we describe the positive impact of ML techniques to the
LoRa [10], one of the most popular LPWAN technologies, problem under consideration and its importance to the future
this trade-off can be tuned by the so-called Spreading Factor 5G standard. The mathematical framework is discussed in
(SF) parameter, which controls the spreading of the signal. Section 4, where a generic RL-based model of nodes is
For larger values of SF, the sensitivity increases (achieving formulated for an arbitrary number of RAT. Furthermore, in
longer transmission distances) whereas the data rate this section, a popular ML genetic algorithm is also proposed
decreases. for deriving the optimal transmission policy. This generic
model is particularized for IoT nodes endowed with multiple
Furthermore, with the aim of reducing cost, most of these RATs in Section 5. Next, in Section 6, the proposed model
LPWAN work in ISM bands. Unfortunately, in many is used to simulate IoT nodes fitted with two RATs: 5G and
countries, these bands are subject to strong regional LPWAN. The ML-derived policy is compared to three other
regulations. For example, in Europe, China and Japan, intuitive policies to further highlight the scope of application
communication devices that do not provide Listen-Before and benefits of our proposal. Finally, Section 7 concludes.
Talk techniques cannot exceed a certain transmission Duty
Cycle (DC) limit [11]. This value is defined as the percentage 2. RELATED WORK
of time that a given device can transmit in a particular
frequency band, usually measured over an hour. In Europe, Nowadays, the wide variety of RATs available on the market
the maximum DC for some bands is limited to 1%, that is, a (cellular networks, LPWAN, Bluetooth, WiFi, ZigBee, etc.)
node cannot occupy the channel for more than 36 seconds makes the idea of adopting a unique and normalized solution
per hour [12]. To accomplish this restriction, after each for IoT, a priori, an unfeasible approach. This is why, in the
transmission, LoRa nodes remain silent during a time period related literature, we can find different reviews and surveys
known as off-period ( ). The duration of such comparing RATs, each with advantages and disadvantages.
follows the expression: = − , where stands In [6], a complete survey about the enabling RATs for the

IoT is conducted, paying special attention to the different
for the duration of the transmission.
LPWAN technologies available for long-range
communications (LoRa, Sigfox, and Ingenu). Following a
As can be seen, there are at least, two competing RATs that similar approach, in [13], an analysis of the potential RATs
cannot, on their own, successfully support the traffic for IoT in a 5G ecosystem is performed. Short-range
generated by the IoT, either due to a lack of specialization in

– 102 –