Decisão do espectro baseado em aprendizado de máquina para redes de rádios cognitivos com suporte a comunicação multi-saltos

Abstract

The technological advances in recent years have reduced the costs of manufacturing wireless devices, increasing the number of applications that use these devices, and most of these applications use the so-called ISM (Industrial, Scientific, and Medical). Therefore, these frequency bands are currently overloaded and thus spectrum occupancy is subject to overlapping problems given the growth in the volume of wireless devices. In order to reduce the interference caused by the large use of spectrum bands that generate information loss and decrease network performance, Cognitive Radios emerge as a smart solution to this problem of spectrum occupancy. Cognitive Radios are software-Defined Radios, which are capable of autonomously performing the reconfiguration of the network by learning and adapting the medium in which it is inserted. For this reason, several types of research were carried out to develop frameworks that could be used in the implementation of Cognitive Radio Networks. However, most researches have developed and evaluated these frameworks using wireless simulators. Although these simulators reflect the environment of a wireless network, the dynamic characteristics of the spectrum are not correctly modeled in the context of Cognitive Radios. Thus, this work proposes the extension of an architecture for Cognitive Radio Networks that aims at the development of methods for spectrum decision based on machine learning algorithms, besides offering the implementation of routing protocols for the network layer to allow multi-hop communication. The developed framework relied on the use of two methods based on machine learning, the first one was the Random Forest that is based on decision trees and the second was the Hidden Markov Model, one stochastic process that uses the Markov property in predicting future unobservable states. In addition, the protocols developed for multi-hop communication were the Ad Hoc On-Demand Distance Vector a reactive protocol, and Optimized Link State Routing a proactive protocol. The results showed that the machine learning methods were able to choose frequency bands of the spectrum in order to decrease the coexistence and thus improve the performance of the network. The method that stands out at this point was the Hidden Markov Model that obtained better results in the package delivery rate.