Page 168 - Proceedings of the 2018 ITU Kaleidoscope

Session 5: Network Applications of Machine Learning
S5.1 Smart Usage of Multiple RAT in IoT-oriented 5G Networks: A Reinforcement Learning Approach
Ruben Martínez Sandoval, Sebastian Canovas-Carrasco, Antonio-Javier Garcia-Sanchez and

Joan Garcia-Haro (Technical University of Cartagena, Spain)
Smart Cities and Smart Industries are the flagships of the future IoT due to their potential to
revolutionize the way in which people live and produce in advanced societies. In these two
scenarios, a robust and ubiquitous communication infrastructure is needed to accommodate the
traffic generated by the 10 billion devices that are expected by the year 2020. Due to its future
world-wide presence, 5G is called to be this enabling technology. However, 5G is not a perfect
solution, thus providing IoT nodes with different Radio Access Technologies (RATs) would allow
them to exploit the various benefits offered by each RAT (such as lower power consumption or
reduced operational costs). By making use of the mathematical framework of Reinforcement
Learning, we have formulated the problem of deciding which RAT should an IoT node employ
when reporting events. These so-called transmission policies maximize a predefined reward
closely related to classical throughput while keeping power consumption and operational costs
below a certain limit. A set of simulations are performed for IoT nodes provided with two RATs:
LoRa and 5G. The results obtained are compared to those achieved under other intuitive policies
to further highlight the benefits of our proposal.

S5.2 Message Collision Identification Approach Using Machine Learning
Juan Pablo Martín, Bruno Marengo, Juan Pablo Prina and Martín Gabriel Riolfo (Universidad
Tecnológica Nacional, Facultad Regional San Nicolás, Argentina)

Machine learning algorithms, in particular k-nearest neighbors (kNN) and support vector machine
(SVM), are employed to estimate the potential success in decodifying ADS-B messages in highly
congested areas. The main aim of this study is to optimize automatic dependent surveillance-
broadcast (ADS-B) reception on-board low Earth orbit satellites. In this first approach, simulations

are performed to obtain the training and testing signals. First, ADS-B communication system is
described; second, machine learning, kNN and SVM are introduced. Third, the developed
simulator is presented and the kNN and SVM algorithms are described with its results. Finally, the
performance of these two is compared.
Optical Flow Based Learning Approach for Abnormal Crowd Activity Detection with Motion
S5.3 Descriptor Map

Dhananjay Kumar and Govinda Raj Sampath Sarala (Anna University, India)

Automated abnormal crowd activity detection with faster execution time has been a major research
issue in recent years. In this work, a novel method for detecting crowd abnormal activities is
proposed which is based on processing of optical flow as motion parameter for machine learning.
The proposed model makes use of magnitude vector which represents motion magnitude of a block
in eight directions divided by a 45 degree pace angle. Further, motion characteristics are processed
using Motion Descriptor Map (MDP), which takes two main parameters namely aggregate
magnitude of motion flow in a block and Euclidean distance between blocks. Here, the angle of
deviation between any two blocks determines which among the eight values in the magnitude
vector to be considered for further processing. The algorithm is tested with two standard datasets
namely UMN and UCSD Datasets. Apart from these the system is also tested with a custom
dataset. On an average, an overall accuracy of 98.08% was obtained during experimentation.

– 152 –