Transfer of Deep Reinforcement Learning for Cloud Service’s Elasticity

• Ano de defesa: 2024
• Autor(a) principal: Cunha, Ian Resende da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Transfer Learning; Deep Reinforcement Learning; Machine Learning; Orchestration; Resource Management; Elasticity; Cloud Service; Cassandra Database; Transferência de Aprendizado; Aprendizado por Reforço Profundo; Aprendizado de Máquina; Orquestração; Gerenciamento de Recursos; Elasticidade; Serviço de Nuvem; Autoscaling; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO; Ciência da Computação; Programas de aprendizado; Dispositivos de treinamento simulado; ODS::ODS 9. Indústria, Inovação e infraestrutura - Construir infraestrutura resiliente, promover a industrialização inclusiva e sustentável, e fomentar a inovação.
• Link de acesso: https://repositorio.ufu.br/handle/123456789/42138; http://doi.org/10.14393/ufu.di.2024.173

Resource management in cloud computing environments is a critical challenge, where an orchestration mechanism seeks to ensure optimized resource utilization while maintaining service quality, preventing waste, and reducing costs. A promising approach to automating this task involves employing machine learning techniques. However, this approach also faces challenges in real-world online training, related to system complexity, extended training durations, safety restrictions, and system rigidity. This dissertation aims to enhance and streamline the Deep Reinforcement Learning training process for tasks related to the orchestration of cloud service resources by employing the Transfer Learning (TL) technique. A source environment was built, comprising a simulation of the target service, and knowledge acquired through simulation training was transferred to enhance training in the real-world service environment. Comparative analysis between TL-based and standard training reveals positive outcomes, including a substantial reduction in time required to achieve reasonable performance, improvements of up to 40% in the initial performance of agents, and up to a 30% enhancement in overall performance during training and testing phases. Finally, it was demonstrated that an agent trained in simulation could be deployed directly into the real environment without additional training, yielding satisfactory and consistent outcomes.