Deep Reinforcement Learning-Assisted NOMA Age-Optimal Power Allocation for S-IoT Network

Abstract

In this paper, we consider a satellite-based Internet of Things (S-IoT) network under shadowed-Rician fading channels, where a satellite transmits timely status updates to multiple user equipments (UEs) with non-orthogonal multiple access (NOMA). In each transmission, the satellite needs to allocate limited power to the status updates for UEs in an appropriate way to guarantee the freshness of updates, characterized by age of information (AoI). To minimize the average AoI of S-IoT network, we formulate a power-constrained optimization problem and then reformulate it as a Markov decision process (MDP). Considering the non-convexity of the optimization problem and the high dimensionality of the multiuser MDP with large state and action spaces, we propose a deep reinforcement learning-assisted age-optimal power allocation (DRAP) scheme to solve the problem and obtain an optimal power allocation policy. Furthermore, a double-network deep reinforcement learning structure is designed to enhance the training effectiveness for our optimization problem. Finally, simulation results show that our proposed DRAP scheme outperforms the benchmark schemes.