Start Date: October 1, 2020
Orfeas Agis Karachalios
The sophistication and complexity of next-generation networks, the wide variety and dynamicity of services anticipated over these networks, and the ever- increasing stringency of service requirements call for the most effective management and orchestration of network resources. Towards this goal, the present research aims at the development of a systematic framework for the coordinated management of multiple and heterogeneous network resources by means of AI/ML techniques. The framework makes use of appropriate graphs to represent topological relationships between relevant resource nodes in the network, but also relationships due to the joint engagement of different resources to provide services. These graphs act as input to modern Deep Learning algorithms that employ Graph Neural Networks (GNNs) to produce a representation of the environment that encapsulates all aforementioned relationships. This representation is exploited as part of the state in the context of modern policy-based multi-agent reinforcement learning (MARL) frameworks, including extensions of the well-established Proximal Policy Optimization algorithm. In the MARL setting, agents represent entities competing for the use of network resources and the actions of the agents jointly contribute to the orchestration of the involved network resources. Employing the GNN-based component in conjunction with the MARL setting is an appealing strategy with high potential for improving the robustness, scalability, modularity, and knowledge transferability attributes of the overall scheme. As a first important domain of application, the research considers the problem of (re)orchestrating resources in the context of network slicing.