Multi-Agent RL based User-Centric Spectrum Allocation Scheme in D2D Enabled Hetnets

Abstract

Device to device (D2D) communication technology is widely considered in 5G for providing higher data rates and increase network capacity. The performance benefits of D2D communication are best achieved if it takes place in shared mode in which it reuses the spectrum being utilized by conventional cellular users. This induces significant challenges in allocating resources because of severe interference among D2D and cellular users. Moreover, centralized resource allocation techniques proposed in literature for D2D users can no longer be practical in dense heterogeneous networks considered for 5G. In this paper, we present a distributed learning based spectrum allocation scheme in which D2D users learn the environment and autonomously select spectrum resources to maximize their Throughput and Spectral Efficiency (SE) while caus- ing minimum interference to the cellular users. We have employed distributed learning in a stochastic geometry based realistic network. Our evaluation results show that the employed learning scheme enables users to achieve high Throughput and Spectral Efficiency while meeting QoS requirements of macro and femto tier.