Two-Sided Learning for NOMA-Based Random Access in IoT Networks

In the Internet-of-Things (IoT), different types of devices can co-exist within a network. For example, there can be cheap but inflexible devices and flexible devices in terms of radio frequency (RF) capabilities. Thus, in order to support different types of devices in different ways and improve throughput, we propose a multichannel random access scheme based on power-domain non-orthogonal multiple access (NOMA), where each flexible or dynamic device (DD) can dynamically choose one of multiple channels when it has a packet to send. In addition, since DDs need to learn the channel selection probabilities to maximize the throughput of DDs, we consider two-sided learning based on a multi-armed bandit (MAB) formulation where rewards are decided by learning outcomes at a base station (BS) to improve learning speed at DDs. Simulation results confirm that two-sided learning can help improve learning speed at DDs and allows the proposed NOMA-based random access approach to achieve near maximum throughput.