SU-Traffic-Aware Deep Reinforcement Learning for Distributed Dynamic Spectrum Access

Spectrum sharing problem of the cognitive radio networks (CRN) with non-stationarity of secondary users (SUs) traffic demands, which could undermine the performances of traditional dynamic spectrum access (DSA) methods, is addressed in this study. The SU traffic arrival pattern is modelled with Poisson process and a SU-traffic-aware multi-agent deep reinforcement learning (MADRL) method is proposed to optimize the spectrum sharing among SUs. Each agent firstly learns the traffic patterns of both local SU and primary users (PUs) to avoid collisions and to improve quality of service (QoS) of SU. Moreover, a priori-rule-assisted DRL algorithm is designed to accelerate agent's training process by enforcing each agent to execute preset actions under specific states. Simulation results show that the proposed method achieves significant improvement on SU QoS while maintains low SU-PU collision probability.