Distributed reinforcement and deep learning for spectrum management in cognitive radio networks
A carregar...
Ficheiros
Data
Autores
Título da revista
ISSN da revista
Título do volume
Editora
Resumo
As redes de rádio cognitivas representam um paradigma promissor para satisfazer a crescente procura de espectro sem fios através do acesso dinâmico ao espectro. No entanto, a deteção eficiente do espectro e a seleção do canal continuam a ser desafiantes devido às características inerentemente dinâmicas e incertas do ambiente sem fios. Esta dissertação apresenta um novo algoritmo distribuído para deteção de espectro e seleção de canais, baseado em mecanismos de consenso em redes rádio cognitivas. A estrutura proposta permite que múltiplos agentes, utilizando a aprendizagem por reforço, aprendam estratégias ótimas para a deteção de espectro, alcançando consenso na seleção de canais de forma descentralizada. A estratégia de consenso facilita a tomada de decisão cooperativa entre agentes numa topologia de comunicação dirigida, variável no tempo e de baixa largura de banda, refletindo condições práticas de implementação descentralizada. Através deste mecanismo, os agentes individuais obtêm conhecimento acionável sobre os estados da rede, permitindo que o sistema atinja coletivamente o desempenho ideal, mesmo quando os agentes isolados possuem informação ou capacidades limitadas. O algoritmo demonstra escalabilidade e robustez a falhas tanto dos agentes como dos enlaces. A dissertação examina as principais características do algoritmo, incluindo os seus efeitos de redução de ruído, a capacidade de coordenação de ações e as melhorias nas taxas de convergência alcançadas através de mecanismos de consenso. A dissertação também investiga a integração de métodos de aprendizagem profunda para melhorar a gestão do espectro em redes sem fios de próxima geração e na Internet das Coisas.
Cognitive radio networks represent a promising paradigm for addressing the increasing demand for wireless spectrum through dynamic spectrum access. Nevertheless, efficient spectrum sensing and channel selection remain challenging due to the inherently dynamic and uncertain characteristics of the wireless environment. This dissertation introduces a novel distributed algorithm for spectrum sensing and channel selection, grounded in consensus-based mechanisms within cognitive radio networks. The proposed framework enables multiple agents employing reinforcement learning to learn optimal strategies for spectrum sensing while achieving consensus on channel selection in a decentralized manner. The consensus strategy facilitates cooperative decision-making among agents over a directed, time-varying and low-bandwidth communication topology, reflecting practical decentralized deployment conditions. Through this mechanism, individual agents obtain actionable knowledge of network states, allowing the system to collectively attain optimal performance even when isolated agents possess limited information or capabilities. The algorithm demonstrates scalability and robustness to both agent and link failures. The dissertation examines key characteristics of the algorithm, including its denoising effects, capacity for coordinating actions, and improvements in convergence rates achieved through consensus mechanisms. The dissertation also investigates the integration of deep learning methods to enhance spectrum management in next-generation wireless and Internet of Things networks.
Cognitive radio networks represent a promising paradigm for addressing the increasing demand for wireless spectrum through dynamic spectrum access. Nevertheless, efficient spectrum sensing and channel selection remain challenging due to the inherently dynamic and uncertain characteristics of the wireless environment. This dissertation introduces a novel distributed algorithm for spectrum sensing and channel selection, grounded in consensus-based mechanisms within cognitive radio networks. The proposed framework enables multiple agents employing reinforcement learning to learn optimal strategies for spectrum sensing while achieving consensus on channel selection in a decentralized manner. The consensus strategy facilitates cooperative decision-making among agents over a directed, time-varying and low-bandwidth communication topology, reflecting practical decentralized deployment conditions. Through this mechanism, individual agents obtain actionable knowledge of network states, allowing the system to collectively attain optimal performance even when isolated agents possess limited information or capabilities. The algorithm demonstrates scalability and robustness to both agent and link failures. The dissertation examines key characteristics of the algorithm, including its denoising effects, capacity for coordinating actions, and improvements in convergence rates achieved through consensus mechanisms. The dissertation also investigates the integration of deep learning methods to enhance spectrum management in next-generation wireless and Internet of Things networks.