Teaching a network to predict bandwidth and steer traffic around congestion, with PyTorch and zero dropped packets.
Networks fail in slow motion far more often than they fail all at once. A link does not usually die. It fills up. Traffic creeps toward the ceiling over minutes, queues build, latency climbs, and by the time an alert fires the users are already unhappy. The interesting question is not “is the link down.” It is “where is this link heading in the next few minutes, and can I do something before it hurts.”
That is a prediction problem, and prediction is what machine learning is for. So in this post we build a closed loop that does five things in order: predict the near future bandwidth of each link, decide whether it is about to cross a limit, steer traffic onto a healthier path if it is, verify that the traffic actually moved, and then restore balance afterward. The whole loop runs on a laptop against a real virtual switching fabric.