Tackling Deployability Challenges in ML-Powered Networks

Following the success of Machine Learning (ML) in various fields such as natural language processing, computer vision and computational biology, there has been a growing interest in incorporating ML into the networking domain [5, 6, 14, 4, 9]. Today, ML-based algorithms for prominent networking problems such as congestion control, resource management and routing, perform very well when their training environment is faithful to the operational environment, achieving state-of-the-art results when compared to traditional algorithms. However, the adaptation of these algorithms to function in production environments has not been straightforward, as real-world networks may differ greatly from the data used for training, leading to a drop in performance when unleashed into the wild.