Dynamically Packet Routing for QoS Assurances on Internet Networks

An adaptive packet routing on the Internet networks based on reinforcement learning has been reported on our empirical study. We show that the approach of Q-routing previously demonstrated on small size of networks which under 100 nodes, can be expanded to large realistic sizes of networks. The performance of such a routing approach on three different network connectivity has been studied: random network, preferential attachment (PA) and a specific architecture known as heuristically optimized topology (HOT). In addition, designed HOT network is related to the Internet's router level topology. Our experimental results show that the HOT network is able to represent significant advantage over a PA network which is characterized by hubs at which communication bottlenecks form.