Page 122 - Proceedings of the 2018 ITU Kaleidoscope

2018 ITU Kaleidoscope Academic Conference

forcing motes endowed with multiple RATs to undergo an
off-period in their LPWAN transceivers, does not have a
measurable effect on their performance. We argue that, if
event-generation rate were dramatically increased (thus,
forcing motes to transmit a larger number of packets), this
forced off-periods may have a non-negligible effect.
However, we also acknowledge that IoT nodes are not
conceived, nor enabled to transmit at such excessive data
rates (mainly due to their limited hardware and energy
resources).

Regarding the results presented in Figure 2, it is worth noting
how the ANN-based proposed policy clearly outperforms the
rest of them when the event-generation rates are larger than
1
Figure 2 – Total reward ( ) as a function of the event- 120 . Specifically, when the proposed policy is preferred over
generation rate. Off-period not enforced. Error bars the second-best policy (5G-first), increases 75.6%, 59.0%,
represent the standard deviation around the mean value. 43.1%, and 17.1% for the four highest event generation rates
respectively. If is smaller or equal to 1 , the relatively low
120
of such policies is a light-weight process that simply entails event-generation rates make the 5G-first policy perform
a forward pass of the ANN (a task computationally viable similarly to our proposal. For these small values of ,
even for very hardware-constrained IoT devices). employing the 5G RAT for all transmissions is always the
best option, since neither the 5G daily quota nor the battery
In order to further show the benefits of the proposed solution, allowance exhaust; thus, both policies attain similar values
the obtained under the proposed solution is compared to of . As can be noted, ANN-based policy is properly adapted
the obtained when three other policies are employed: to different event-generation rates, always achieving an
optimal performance. Unarguably, these results are subject
(i) randomly chose a RAT (5G/LoRa) –denoted as random to change when the 5G daily quota or the battery capacity
policy–, (ii) start with 5G until its exhaustion and then use changes; however, we believe that the main idea is
LoRa if necessary –denoted as 5G-first policy–, and (iii) adequately pinpointed: as nodes transmit more and more data,
consider important events those with > 0.5 and thus, only long-lasting technologies such as LoRa (both in terms of
employ 5G for those critical transmissions (and employ extending lifespans of the nodes and in their usage limitation)
LoRa when ≤ 0.5 or when 5G has been exhausted) – should be progressively embraced. Another interesting fact
denoted as priority-based policy–. should be highlighted: if non-adaptive policies (such as 5G-
first, priority-based, or random) are applied, and the event-
Figure 2 depicts the obtained results for different values of generation rates are relatively high ( to ), it is the battery
1
1
. It is worth mentioning that, values obtained when an off- 30 60
period is enforced after every transmission are remarkably capacity the limiting factor in attaining larger values of .
close to those obtained when such limitation is not These rates make batteries deplete half-way through the 24-
considered (differences are, in average, less than 0.047%). hour simulation, and thus, larger values of do not lead to
Therefore, due to space limitations, we have only included larger values of for the three aforementioned policies. This
the results obtained when an off-period is not enforced. reveals that, when designing transmission policies for
These small differences indicate that, for typical IoT event- hardware-constrained IoT devices with multiple RAT, it is
generation rates (i.e. those evaluated in this simulation), the battery consumption, and not the bitrate of their

Figure 3 – RAT usage for the two best policies: the proposed one (left) and the 5G-first (right) for 1000 packet
transmissions. Obtained for = ⁄ and off-period not enforced.
1
30

– 106 –