ORCA Seminar: Multi-Agent Deep Reinforcement Learning for Request Dispatching in Software-Defined Networks

26 Jan 2021 12:00 PM - 1:00 PM
Location: https://meet.google.com/dvy-vtkp-cvv

Recently, distributed controller architectures have been quickly gaining popularity in Software-Defined Network (SDN). However, the use of distributed controllers introduces a new and important Request Dispatching Problem with the goal for every SDN switch to properly dispatch their requests among all controllers so as to optimize network performance. This goal can be fulfilled by designing a dispatching policy to guide request distribution at each switch. Existing policies are designed under the assumption of a single agent and fully observable environment which cannot always be satisfied in practice. In this talk, we present a Multi-Agent Deep Reinforcement Learning (MA-DRL) approach to automatically design policies with high adaptability and performance. This is achieved through a new problem formulation in the form of a Multi-Agent Markov Decision Process, a new adaptive policy design, and a new MA-DRL algorithm. Extensive simulation studies show that our MA-DRL technique can effectively train policies to significantly outperform man-made policies, model-based policies, as well as policies learned via single-agent DRL algorithms.