LAUSR

Packet-level forward erasure correction (FEC) is effective for achieving low-latency transmissions in non-terrestrial networks (NTNs), which often contain lossy links with long propagation delay. This letter considers the problem of dynamically sending FEC coded packets based on delayed feedback to achieve fast in-order data delivery without retransmissions. We show that the problem can be formulated using reinforcement learning (RL) by tracking history feedback and system state. With carefully devised state space and reward signal, the RL scheme is shown to be effective for achieving high and smooth in-order goodput with low delay. For example, smooth goodput can be achieved at a FEC code rate of more than 90% of the capacity with the delay of all the packets being lower than 1.5 round-trip time, which would be incurred by a single packet loss if using retransmissions.