IEEE Consumer Electronics Magazine - January/February 2023 - 38

Autonomous UAV Base Stations Through Deep Learning
Figure 3. Locations of the UAV-BS for the optimal and CNN-based solutions. Light blue disks and inverse
triangles represent the initial and final locations of MUs, respectively. Dark blue dashes show the trajectory of
the users. Red crosses and black diamonds indicate the locations obtained by the optimization algorithm and
the proposed CNN approach, respectively. Yellow disks signify the proximity of the estimated locations.
various environment types. We utilize the parameter
set for the urban environment. If the objective
is to maximize the information rate then it is
necessary to consider the effects of instantaneous
fading and noise level20. However, for coverage
maximization, utilization of average path
loss is the common approach3,8,10,11,21 which is
also the approach employed in this study. The
rest of the parameters are adopted from Bor-Yaliniz
et al.'s work.10
For each instance of the scenario, we compute
the best UAV-BS position, which will be used as a
label during the learning stage, using (1) by considering
only the coverage objective and without
considering the temporal dimension. To this end,
we utilize the UAV-BS placement approach proposed
in Bor-Yaliniz et al.'s work.10
In Figure 3, UAV-BS locations given by our
model and the optimization algorithm are presented
for three test cases. As Figure 3 clearly
shows the UAV-BS locations estimated by the
CNN-based approach are in close proximity of the
optimal locations. To investigate the performance
of the CNN-based solution in comparison to RLbased
approaches, we use three performancemetrics.
First, to quantify the difference between the
estimated location of the UAV-BS and the optimal
location, we utilize MSE and MAE metrics, which
are generally used to evaluate the performance of
the regression-based learning approaches.18 We
employ the cumulative distributive function (CDF)
of the covered MUs as the second metric to statistically
analyze the coverage performance of the
38
learning approaches. Third, as a communication
performance indicator, we use the average number
of coveredMUs for a particular time sequence.
Hence, we can examine the success of different
learningmodels for dynamic communication environment.
We obtain results with DRL, Q-learning,
and double Q-learning4,7,12 for these three metrics.
Figure 4 illustrates the block diagram of the RL
approaches, where CRL indicates the conventional
RL approaches (i.e., Q-learning and double
Q-learning). The parameters utilized to realize the
RL approaches are also provided in Figure 4. Furthermore,
the discount factor for DRL is 0.99, and
the loss function to update the weights of the DNN
is theMSE.
MAE and MSE scores between the optimal
solution and those obtained for ML-based
approaches are provided in Table 1, which reveals
Figure 4. RL approaches and their parameters.
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - January/February 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2023
