A Study of a Competitive Reinforcement Learning Approach for Joint Spatial Division and Multiplexing in Massive MIMO

Abstract

Massive Multiple-Input Multiple-Output (MIMO) is widely considered a pivotal communication technology for future generations of wireless networks. Massive MIMO uses a large number of antennas at the base station, which offers better effectiveness in spectral and energy use. However, a Frequency Division Duplex (FDD) system is challenging in reciprocity since it is difficult to estimate channels and requires feeding back channel state information. Joint Spatial Division and Multiplexing (JSDM) is a simplified FDD technique to provide massive MIMO gains. The main idea of JSDM is related to grouping users with approximately similar channel covariance. Many machine learning algorithms have been applied to conduct user grouping. In this paper, to improve the user grouping, we employ Reinforcement Guided Competitive Learning (RGCL) to the user grouping and then compare it with clustering techniques, including K-means, and sequential K-means to achieve the appropriate user grouping. The experimental results show that the RGCL technique represents better performance in computational time and system throughput than the other two above mentioned techniques, since RGCL can avoid being trapping in local minima.