Real-Time TCP Packet Loss Prediction Using Machine Learning

Congestion and resulting packet loss in TCP connections can lead to performance degradation and reduce the Quality of Experience (QoE) for end users. Many common TCP congestion control algorithms therefore adjust their sending rate proactively, using heuristics based on measured delay and/or throughput. We have investigated whether it would be possible to replace such heuristics with Machine Learning (ML), such that Internet hosts could meaningfully react when a packet loss is predicted to happen. For example, such a reaction could involve the use of Forward Erasure Correction (FEC) to protect against the loss, or to dynamically tune parameters of TCP. We present two ML models that were trained for Reno and Cubic flows, respectively, and show that they can be beneficially applied in real time. We do this by informing a TCP sender to reduce its rate via a local, artificially generated Explicit Congestion Notification (ECN) signal at the sender before packet loss would have happened.