Network Routing Optimization Based on Machine Learning Using Graph Networks Robust against Topology Change
Kaku Sawada, Daisuke Kotani, Yasuo Okabe · 2020
There is an increasing demand of real-time routing optimazation using Sotware Defined Networking (SDN) for better Quality of Service. Since the problem of finding the optimum routing for any QoS metric is hard to solve for a medium or larger size network, quasi-optimization using metaheuristics, such as Genetic Algorithm (GA) and Simulated Annealing (SA), have been investigated, but it is still impossible to satisfy the requirement of real-time optimization. There have been some attempts to solve this with machine learning. By learning a model beforehand, it is possible to output a near-optimal solution in a short time during network operation. The open problem with this approach is that machine learning models cannot deal with topology change of the network. In this paper, we create a model which is robust for topology change by using Graph Networks. Applying the proposed model to maximum bandwidth utilization, we have gotten the accuracy of about 61.0% for solution of GA, and the prediction time is 150 times faster than GA.