Deep-Reinforcement Learning for Fair Distributed Dynamic Spectrum Access in Wireless Networks

Many studies investigated fair dynamic spectrum access in distributed wireless networks (DWNs). To limit the required communication between nodes, several schemes based on reinforcement learning (RL) have been proposed and their results are promising. One missing point in these studies is that they assume that all users are in saturated mode (always have data to transmit) which is not a correct assumption in reality. In this paper, we remove this assumption and propose a multi-agent RL scheme that incorporates the current need of a user to transmit (its buffer level) in deciding if a user should transmit or not. The complexity though each user is not aware of the buffer level of other users except some delayed information that it can receive from the previous behavior of those users. As simulation results show, the proposed method maximizes the total throughput while trying to make fair resource allocation by first serving the user with the highest level of packets in its queue.