Page 123 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

transceivers, the main limiting factor in increasing network hardware-constrained IoT devices. The solution based on an
performance. By carefully discarding non-critical events in Artificial Neural Network (ANN) optimizes the use of the
more demanding scenarios (i.e. where larger values of are available RATs (5G and LoRa), leading to an improvement
present), our proposed policy can get the most out of the in the attained rewards of up to 75.6% when compared to
allocated resources (battery and 5G daily quota) to maximize other policies. Although the precise presented results depend
obtained rewards. Figure 3 corroborates this statement by on the parameters of the simulated network (and, deeply
analyzing the typical usage of each RAT made by our studying other scenarios is left as a future work), we truly
proposed policy and the 5G-first policy. 5G transmissions believe that this work has served its purpose: to raise
are depicted in blue stars, LoRa transmissions in red circles, awareness about the potential use of ML techniques in
and discarded packets in black crosses. Two variables are deriving transmission policies that could make the future 5G
analyzed: packet length and packet priority (the rest of standard feasible for the IoT revolution.
variables are not shown for the sake of clarity, although they
still play an important role in deciding on the RAT). Note Furthermore, there are plans to extend the presented work
how the proposed policy (left picture) tends to discard large with traffic traces generated from IoT devices –so that the
non-important packets (shown in the bottom right side of the derived ANN-based transmission policy can learn from real
picture) while high-importance packets are normally data–. Nevertheless, authors believe that the traffic
transmitted via 5G-cellular links. LoRa is employed when generation patterns are in-line with actual IoT deployments
either the packet priority is not high, the packet is relatively (as indicated in the Simulation Section).
large or the 5G daily quota has been exhausted (this variable
is not shown in Figure 3). Conversely, the 5G-first policy 8. REFERENCES
disregards packet properties and makes an excessive early
use of 5G transmission (note that there are far more blue stars [1] P. Kulkarni and T. Farnham, “Smart City Wireless
than red circles) until 5G daily quota is exhausted and LoRa Connectivity Considerations and Cost Analysis:
comes into play. Discarded packets are not prioritized by Lessons Learnt From Smart Water Case Studies,”
their importance and thus, there is a homogenous distribution IEEE Access, vol. 4, pp. 660–672, 2016.
of black dots.
[2] R. Martinez-Sandoval, A. J. Garcia-Sanchez, F.
The reason why the priority-based policy performs worse Garcia-Sanchez, J. Garcia-Haro, and D. Flynn, “A
than the 5G-first policy lies in the usage of the 5G RAT. The comprehensive WSN-based approach to efficiently
former tends to systematically underuse 5G, whose daily manage a smart grid,” Sensors, vol. 14, no. 10, pp.
quota is not met even for the lowest values of (in such 18748–18783, Jan. 2014.
scenarios, it is the battery and not the 5G daily quota the
limiting factor). Finally, the random policy makes the same [3] F. Orfei, C. Benedetta Mezzetti, and F. Cottone,
mistakes as the priority-based policy while also disregarding “Vibrations powered LoRa sensor: An
event priorities; hence the lower attained . electromechanical energy harvester working on a
real bridge,” in 2016 IEEE SENSORS, 2016, pp. 1–
7. CONCLUSION 3.
[4] J. G. James and S. Nair, “Efficient, real-time
5G has proved itself to be a future revolution in the way in
which, not only users, but also machines could communicate tracking of public transport, using LoRaWAN and
RF transceivers,” in TENCON 2017 - 2017 IEEE
with their peers. It is precisely in this type of
communications (machine-to-machine) in which the IoT Region 10 Conference, 2017, vol. 2017–Decem, pp.
2258–2261.
paradigm will play a major role. However, cellular networks
may not be fully geared towards the very sporadic and bursty [5] A. Orsino, G. Araniti, L. Militano, J. Alonso-Zarate,
nature of IoT data transfers, and many authors have argued A. Molinaro, and A. Iera, “Energy Efficient IoT Data
that IoT devices could potentially benefit from having Collection in Smart Cities Exploiting D2D
multiple Radio Access Technologies (RAT). These RATs Communications,” Sensors, vol. 16, no. 6, p. 836,
would be used intelligently and on-demand to access the Jun. 2016.
wireless shared medium.
[6] M. Centenaro, L. Vangelista, A. Zanella, and M.
To evidence this, we have presented a mathematical model Zorzi, “Long-range communications in unlicensed
that lies the foundations for improving the performance of bands: The rising stars in the IoT and smart city
IoT devices fitted with multiple RAT. To optimize such a scenarios,” IEEE Wirel. Commun., vol. 23, no. 5, pp.
figure of merit, we have resorted to a set of analytical tools 60–67, Oct. 2016.
popularized by the Machine Learning (ML) research field. In
particular, a Genetic Algorithm known as Evolution [7] M. Agiwal, A. Roy, and N. Saxena, “Next
Strategies has been used to derive optimal transmission Generation 5G Wireless Networks: A
policies. These policies are shown to outperform other Comprehensive Survey,” IEEE Commun. Surv.
intuitive approaches in different simulated scenarios while Tutorials, vol. 18, no. 3, pp. 1617–1655, 2016.
still being computationally light-weight enough to run in

– 107 –