AI gym for Networks

• Autor(a) principal: Capelo, Pedro Alexandre Francisco
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/152022
• topic: 5G Networks; Deep Reinforcement Learning; Containernet,; Ryu Controller; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

5G Networks are delivering better services and connecting more devices, but at the same time are becoming more complex. Problems like resource management and control optimization are increasingly dynamic and difficult to model making it very hard to use traditional model-based optimization techniques. Artificial Intelligence (AI) explores techniques such as Deep Reinforcement Learning (DRL), which uses the interaction between the agent and the environment to learn what action to take to obtain the best possible result. Researchers usually need to create and develop a simulation environment for their scenario of interest to be able to experiment with DRL algorithms. This takes a large amount of time from the research process, while the lack of a common environment makes it difficult to compare algorithms. The proposed solution aims to fill this gap by creating a tool that facilitates the setting up of DRL training environments for network scenarios. The developed tool uses three open source software, the Containernet to simulate the connections between devices, the Ryu Controller as the Software Defined Network Controller, and OpenAI Gym which is responsible for setting up the communication between the environment and the DRL agent. With the project developed during the thesis, the users will be capable of creating more scenarios in a short period, opening space to set up different environments, solving various problems as well as providing a common environment where other Agents can be compared. The developed software is used to compare the performance of several DRL agents in two different network control problems: routing and network slice admission control. A novel DRL based solution is used in the case of network slice admission control that jointly optimizes the admission and the placement of traffic of a network slice in the physical resources.