Adaptive Configuration of a Queuing Policy on a Switch, Obtained via Machine Learning

The problem of traffic balancing is relevant in modern networks that have many alternative routes between any pair of subscribers. Balancing allows us to utilize network resources uniformly. In this work, we propose a way of adaptively tuning the queuing policy on a switch to achieve uniform queue utilization on the output ports of the switch. Since modern applications limit the transmission delay to milliseconds, a machine learning technique with DQN reinforcement learning is used to solve the problem. Experimental study shows the convergence of the proposed approach during the learning process to uniform queue utilization at the output ports.