Hierarchical Multi-Objective Deep Reinforcement Learning for Packet Duplication in Multi-Connectivity for URLLC

In this paper, machine learning solutions have been investigated to improve the decision of packet duplication in a multi-connectivity cellular network to optimize the satisfaction of delay and reliability in 5G. A multi-agent deep reinforcement learning scheme with sequential actor-critic model has been developed to improve the decision of packet duplication from observations of radio environment including channel state, interference and load. A multi-objective reward function has been developed to minimize the transmission delay, error rate and maximize satisfaction of the URLLC targets. System-level simulations have been carried out in a heterogeneous network by utilizing dual connectivity between macro and small cells. Our deep reinforcement learning scheme is shown to prioritize packet duplication to the UE where it gains from lower queueing and interference. Comparing with standard 5G multi-connectivity, it reduces the overall packet error rate and delay, with increased satisfaction rate of URLLC targets. Furthermore, it improves the network throughput and resource efficiency in dynamic user traffic with lower redundancy.