Graph Neural Networks for Timely Updates of Short Packets in Interference-Limited Networks

This paper investigates the timely updates of short packets in interference-limited networks. Specifically, we minimize a performance metric, Age-of-Information (AoI), by optimizing the power control policy that is scalable to the number of wireless links. Since the AoI minimization problem in the interference-limited network is a sequential decision-making problem, we develop a deep reinforcement learning algorithm, graph-based deterministic policy gradient (GDPG), to find the optimal policy. The actor and critic of the GDPG are represented by two types of graph neural networks (GNNs). The actor is a random edge GNN that maps the AoI and network topology to the transmit power. The critic is a message-passing GNN for estimating the long-term reward of state-action pairs visited by the policy, where the received interference is the aggregated message of a node in the message-passing GNN. Simulation results show that even when the number of wireless links in the testing stage is different from that in the training stage, the average AoI achieved by the GDPG is 30% lower than the weighted minimum mean square error (WMMSE) algorithm. Therefore, the proposed GDPG is scalable to the number of wireless links in the communication network.