Deep Reinforcement Learning Based Intelligent User Selection in Massive MIMO Underlay Cognitive Radios

Cognitive radio (CR) and massive multiple-input multiple-output (MIMO) have attracted much interest recently due to the amazing ability to accommodate more users and improve spectrum utilization. This paper investigates the QoS-aware user selection approach for massive MIMO underlay cognitive radio. Two main CR scenarios are considered: 1) the channel state information (CSI) of the cross channels are available at the secondary base station (SBS), and 2) any CSI of cross-network is unavailable at SBS. For the former, we develop the low-complexity increase-user-with-minimum-power algorithm (IUMP) and decrease-user-with-maximum-power algorithm (DUMP) which both can address the problem of user selection with power allocation. However, the CSI is typically not available in practice. To address the intractable issue, we propose a deep reinforcement learning-based approach, which can enable the SBS to realize efficient and intelligent user selection. The simulation results show that the IUMP and DUMP algorithms have obvious performance advantages over traditional user selection methods. In addition, results also verify that our constructed neural network can efficiently learn the optimal user selection policy in the unknown dynamic environment with fast convergence and high success rate.