Soluções baseadas em aprendizado de máquina para alocação de recursos em redes sem fio de próxima geração

Abstract

5G and beyond networks have been designed to support challenging services. Despite important advances already introduced, resource allocation and management methods remain critical tasks in this context. Although resource allocation methods based on exact optimization have a long history in wireless networks, several aspects involved in this evolution require approaches that can overcome the existing limitations. Recent research has shown the potential of AI/ML-based resource allocation methods. In this approach, resource allocation strategies can be built based on learning, in which the complex relationships of these problems can be learned through the experience of agents interacting with the environment. In this context, this thesis aimed to investigate AI/MLbased approaches for the development of dynamic resource allocation and management methods. Two relevant problems were considered, the rst one related to user scheduling and the allocation of radio resources in multiband MIMO networks, and the second one focused on the challenges of allocating radio, computational, and infrastructure resources involved in the VNF placement problem in disaggregated vRAN. For the rst problem, an agent based on DRL was proposed. For the second problem, two approaches were proposed, the rst one being based on an exact optimization method for dening the VNF placement solution, and the second one based on a DRL agent for the same purpose. Moreover, components adhering to the O-RAN architecture were proposed, creating the necessary control for monitoring and dening new placement solutions dynamically, considering aspects of cell coverage and demand. Simulations demonstrated the feasibility of the proposals, with important improvements observed in different metrics.