IEEE Consumer Electronics Magazine - January/February 2023 - 37

Second, the model is resilient to the ingress/egress
of MUs to/from the area under consideration.
Third, the model facilitates the use of only a subset
of theMUs in theinput,which canbe used to
decrease the computational complexity. Although
spatial and temporal clustering of the MUs can
decrease precision, it increases the ability to generalize,
which improves the robustness of the model.
CNNs can exploit spatiotemporal characteristics of
data more efficiently when compared to other
learning approaches due to the embedding of various
convolutional filters within the CNN architecture.
The duration of the temporal horizon can be
adjusted according to the operational dynamics
and the environment. However, increasing the
duration of the temporal horizon inflates both the
training time and memory requirement. Nevertheless,
overextending the temporal horizon would
lead to non-negligible deterioration of precision,
whichshouldbeavoided.
The most computationally demanding part of
the proposed approach is the training stage,
which is performed offline. Once the model has
been trained sufficiently, the solution can be
obtained almost instantly. Since the CNN-based
approach has a holistic view of the problem, the
UAV-BS trajectory is created by considering multiple
time instances simultaneously. The proposed
model is resilient to varying channel characteristics
and movements of MUs (i.e., once the CNN is
trained satisfactorily, the trajectory is computed
effortlessly for any input vector without the need
to rerun an optimization algorithm).
USE CASE SCENARIO
We generate a UAV-BS deployment scenario
to analyze the performance of the CNN-based
solution. To train the CNN, we create a synthetic
dataset. First, we obtain the optimal locations of
the UAV-BS for a range of MU deployments using
reference model parameters.10 Next, we use the
constraints and locations of MUs as inputs and
optimal locations of the UAV-BS as the output to
train our CNN model. Finally, we test the CNN
model to compute the locations of the UAV-BS
for new user deployments (i.e., the test cases).
Moreover, we provide comparisons of the solutions
obtained with the CNN approach and various
RL approaches.
January/February 2023
We formulate the problem as a regression
problem, for which MAE is the standard metric.18
We are aware that it is not the only way to
formulate the UAV localization problem as one
can discretize the solution space using nonoverlapping
grids, which translates the problem into
a classification task (i.e., predicting the grid
point, where the UAV-BS resides). However, this
only provides a broader approximation over the
regression-based solution (i.e., as grids get arbitrarily
finer the UAV location can be represented
on continuous space). Thus, formulating the
UAV-BS positioning problem as a regression
problem yields better precision in determining
the coordinates of the UAV-BS.
We demonstrate the efficiency of the proposed
CNN-based approach by using a typical
urban environment communications scenario
served by a UAV-BS. 30 MUs are distributed over
a 2 km by 2 km area. The locations of MUs are
within the radius of a predetermined center,
which is found using a uniform random distribution.
Each MU moves toward a randomly chosen
direction with a constant speed (i.e., randomwaypoint
mobility model19). A series of 15 consecutive
moves of an MU is called a session. MUs
restart their movements at random positions at
the beginning of each session.
Our scenario consists of 90,000 session instances.
Since we have a substantial amount of training
samples, we train our network with a single
train/test split instead of conducting k-fold crossvalidation.
To have a satisfactory variation within
the test set, we utilize a rather generous train/test
split size of 80/20, which yields a high amount of
test samples to evaluate our model, thus increasing
our confidence in the final results. The number
of sessions for training and test stages are
72,000 and 18,000, respectively.
Although it is possible to consider a fixed
deployment topology and compute the path loss
through raytracing, as is done in Bayerlein et al.'s
work20, we adopt the path loss model employed
in Bor-Yaliniz et al.'s work10, which is a widely utilized
air-to-ground path loss model originally proposed
in Al-Hourani et al.'s work.21 This model
considers both LoS and non-LoS probabilities
depending on the vertical and horizontal distances
between the UAV-BS and MUs. The parameters
of the model can be adjusted to model
37
IEEE Consumer Electronics Magazine - January/February 2023

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