Traffic Restoration in Communication Networks by Meta-Learning Inspired Algorithm Selection: A Case Study for IP-Optical SDN Networks

Abstract

The performance of optimization algorithms for traffic restoration in communication networks is dependent on the characteristics of the problem instance. There is no known best algorithm that performs well for all instance realizations of any given traffic restoration problem. In this paper we argue that for a given problem instance, optimum or near-optimum performance can be attained through algorithm selection (AS). The objective is to select from a set of candidate algorithms, the one that performs best for the problem instance. For that, AS is formulated as a learning problem where a Machine-learning algorithm learns the relation between the instance properties and the performance of the candidate algorithms. As case study, the approach is applied for traffic restoration in IP-Optical networks. In these networks, optical failures may affect multiple IP links simultaneously. As a result, the IP layer has to perform traffic restoration by rerouting the affected IP flows. This can be carried out by a hyperheuristic method that performs traffic engineering to minimize the spare capacity utilized for traffic protection. The approach applies AS to choose the best algorithm from a set of heuristics for IP traffic rerouting. Results on selected scenarios show that AS predicts with high precision the heuristic that requires minimum spare capacity.