Abstract:
The 5G technology has created a lot of interest recently to support the increasing
demand for higher data rates by the user equipments (UEs). The most interesting
part of 5G is the use of device to device (D2D) communication, using millimeter
waves. Millimetre waves can immensely increase the data rates compared to 4G. But
millimetre waves su er from a host of problems ranging from high free space loss
restricting range, and also extremely high penetration loss, making it almost line of
sight (LOS) communication. To handle these problems, communication links can be
broken into multiple hops by using intermediate devices as relays, avoiding obstacles
and thereby extending the range. Knowing the location and size of obstacles is key to
choosing good relays. Satellite imagery can be used to do the same. But small obsta-
cles like trees cannot be captured properly using satellite imagery. Moreover presence
of an obstacle in the image, does not guarantee obstruction. Our work concentrates
on mapping static obstacles in a given area to help in the process of relay selection.
We propose a learning based strategy, which considers spatial correlation of obstacles,
to build the static obstacle map e ciently without relying on satellite imagery. We
propose a evidential framework to model the con dence in the knowledge gained for
each cell, as this can model uncertainty better than a typical probabilistic model. We
also propose an operation similar to Gaussian smoothing, to learn information about
a cell, using the information available from the nearby cells. We have proposed a
visibility graph algorithm which will reduce the overhead due to repeated updation of
the location information by the UEs, and thereby improve the overall coverage dra-
matically. We have also proposed a exploration mode during relay selection, which
can be used in the place of the normal mode of relay selection, which can reduce
the time taken to learn the map, by selecting relays which encourage exploration
rather than prioritizing maximizing throughput only. Through simulations, we show
that our evidential framework based approach can learn the map faster and more
accurately than the typical probabilistic model based work [18].