Network traffic feature representation with contrastive learning for traffic engineering in hybrid software defined networks

Abstract

Traffic Engineering (TE) promotes the performance of hybrid Software Defined Networks (hybrid SDN) through optimizing traffic route selection. To handle dynamic network traffic, existing machine learning-based TE methods in hybrid SDNs focus on leveraging Reinforcement Learning (RL) to learn the mapping between the dynamic traffic demands and the traffic splitting ratios. However, with the huge network state space incurred by the dynamic network traffic and increasing network scale, it is hard for the RL-agent to learn and converge to the optimal mapping between traffic demands and traffic splitting ratios, thus the network performance suffers a degradation in dynamic network environment. To tackle this issue, we innovatively propose a TE approach that combines Contrastive learning (CL) and RL. Specifically, to reduce huge state space, we design to learn the mapping between network traffic features and routing policy rather than learning the mapping between traffic demand and routing policy. To well capture the features of traffic demands, we leverage CL to train a feature encoder for representing network traffic. We conduct extensive experiments on real network topologies datasets and the experimental results demonstrate that our proposed algorithm provides significant network performance improvements over state-of-arts.