Network Link Weight Setting: A Machine Learning Based Approach

Internet routing protocols like OSPF and ISIS use shortest path routing to route traffic from ingress nodes to egress nodes in a network. These shortest paths are computed with respect to the weights assigned to links in the underlying network. Since the routed paths depend on the assigned link weights, a fundamental problem in optimizing network routing is the determination of the set of weights that minimizes congestion in the network. This is an NP-hard combinatorial optimization problem. Consequently, several heuristics have been developed to determine the set of link weights to minimize congestion. In this paper, we develop a machine-learning based approach by formulating a smoothed version of the weight setting problem and using gradient descent in the PyTorch framework to derive approximate solutions to this problem. We demonstrate the improvement in performance compared to traditional approaches using several benchmark network topologies.