Troca automática de protocolos MAC empregando aprendizado por reforço

• Ano de defesa: 2019
• Autor(a) principal: André Vinicius Gomes Santos Gonçalves
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: seleção protocolos MAC; Aprendizado por reforço; Aprendizado de máquina; subcamada MAC; Redes; Computação; Redes de Computadores; Aprendizado do computador; Sistemas de comunicação sem fio
• Link de acesso: http://hdl.handle.net/1843/ESBF-BB2J75

There is an increasing demand for wireless networks propelled by the popularization of mobile devices such as laptops, smartphones and tablets, and by emerging technologies such as the internet of things. If the demand is increasing on one hand, the efficient use of wireless networks is challenging on the other. For instance, wireless networks are at mercy of changes in the network topology, obstacles between devices, node's mobility, propagation medium fluctuations such as humidity or weather conditions etc. Further, the same medium is used by all devices, resulting in competition between nodes. Despite the increasing demand, dynamism and complexity of wireless networks, some communication protocols focus on specific application scenarios, which is the case for MAC protocols. For example, contention-based MAC protocols tend to perform better in networks under low competition, becoming inefficient as competition rises. Conversely, contention-free MAC protocols tend to perform better when network competition is high. Since wireless networks are dynamic environments, a single MAC protocol is unlikely the best one all the time. A more efficient approach is to switch the MAC protocol according to the network conditions over time. This dissertation tackles that problem, consisting of a Self-Organizing MAC sublayer (SOMAC) capable of automatically selecting and switching the MAC protocol over time. This solution is feasible for two reasons. The first is called software defined radios, consisting of programmable radios that allow the flexibilization of the MAC sublayer. The second corresponds to machine learning techniques, allowing an automatic MAC protocol selection. SOMAC deploys both: a) implementation and validation through software defined radios and b) a selection engine powered by machine learning. SOMAC uses the reinforcement learning algorithm Q-Learning and copes with the dynamic nature of wireless networks. Our solution periodically assesses the wireless network performance through network metrics and selects the best MAC protocol. In addition, SOMAC is implemented and validated in real-world wireless networks. Our results indicate that SOMAC selects the best MAC protocol at a minimum rate of 80% of the time, reaching up to 90% of optimality. Furthermore, they also point that SOMAC defeats its main competitors by a significant extent, outperforming the state of the art.