Reinforcement learning-based dynamic bandwidth provisioning for quality of service in differentiated services networks

The issue of bandwidth provisioning for Per Hop Behavior (PHB) aggregates in Differentiated Services (DiffServ) networks is imperative for differentiated QoS to be achieved. This paper proposes an adaptive provisioning scheme that determines at regular intervals the amount of bandwidth to provision for each PHB aggregate, based on traffic conditions and feedback received about the extent to which QoS is being met. The scheme adjusts parameters to minimize a penalty function that is based on the QoS requirements agreed upon in the service level agreement (SLA). The novel use of a continuous-space, gradient-descent reinforcement learning algorithm enables the scheme to work effectively without accurate traffic characterization or any assumption about the network model. Using ns-2 simulations, we show that the algorithm is able to converge to a policy that provisions bandwidth such that QoS requirements are satisfied.