A collaborative and distributed learning-based solution to manage computer networks

The high programmability provided by Software Defined Networking (SDN) paradigm facilitated the integration of Machine Learning (ML) methods to design a new family of network management schema. Among them, we can cite self-driving networks, where ML is used to analyze data and define strategies that are then translated by SDN controllers into network configurations, thus making networks autonomous and capable of auto-scaling decisions based on the network's needs. Despite its attractiveness, the centralized design of the majority of proposed solutions cannot keep up with the increasing size of the network. To this end, this thesis investigates the use of a multi-agent reinforcement learning (MARL) model for auto-scaling decisions in an SDN environment. Physically distributed controllers raise new challenges to be addressed: in absence of a central entity with global knowledge, controllers must cooperate sharing their information in order to make the best decision. In particular, we study two possible alternatives: a logically decentralized control plane and a fully distributed control plane. The former implementation is defined by multiple controllers that collect statistics over their own area and share them to feed the model with the same input; the latter deploys many controllers that collaborate to correctly perform routing but feed the model with their local information only. Results showed that both approaches can guarantee high throughput while minimizing the set of active resources.