Application of reinforcement learning for the control of packet routers

Abstract

With increasing importance, Internet-based applications need of a more and more complex mesh of networking infrastructure to address the stringent connectivity requirements that they require. 5G networking is being developed to support the needs to such applications. No 5G applications can be supported without an underlying 5G infrastructure. However, the control and management of such complex data plane also requires from smart and advanced solutions to make services affordable in terms of cost for application and infrastructure owners. As enabler of this smart control and management concept, a plethora of artificial intelligence (AI) and machine learning(ML)-based solutions have been recently proposed. Nevertheless, training such models could be a hard (even unfeasible) task and, recently, reinforcement learning (RL) solutions are being proposed to solve challenging networking problems. In this project, we aim at designing and developing a RL-based methodology for smart management of network equipment. In particular, we focus on creating a module to be locally deployed in a packet node(e.g. router) to autonomously adjust interface buffer to actual traffic needs.