Towards a tight and fast interaction of data and control planes

With the birth of Software Defined Networks (SDN) a new impulse was given to networks innovation, making it possible to decouple data plane and control plane. Several technologies were born with the aim of speeding up the implementation of new features, exclusive task of hardware manifacturers for years. One of them is P4, a technology capable of enabling the general-purpose programming paradigm in the networking world, thus telling the devices what exactly to do with incoming packets. Thanks to this new technology (2016), this thesis aims at proposing a new concept of SDN that breaks the main pillar of having one logical centralized control plane. Each network device has a limited view on the topology and will learn how to route packets thanks to a Deep Reinforcement Learning training, performed in parallel with all the other devices of the network. This learning technique combines Reinforcement Learning (RL), i.e. learning by trial and error basing on feedbacks of the environment to the chosen actions, and Deep Learning, a branch of Machine Learning which can be used to enable the adoption of RL algorithms for problems whose state space is high-dimensional.