An Unsupervised Learning Paradigm for User Scheduling in Large Scale Multi-Antenna Systems

The tremendous growth of mobile networking and Internet of Things (IoT) demands efficient and reliable service for massive wireless systems. Multi-input-multi-output (MIMO) technologies successfully utilize spatial diversity to substantially improve spectral efficiency by scheduling multiple devices for simultaneous spectrum access. Efficient solutions to the NP-hard problem of scheduling large number of users are vital to interference mitigation and spectrum efficiency. Despite successes of machine learning in tackling large-scale optimization problems, direct adoption of supervised learning in MIMO user scheduling is difficult as there is no optimum solution to use as labeled training data, and unsupervised learning would identify similar user channel features instead of promoting channel diversity. In this work, we propose an effective and scalable user scheduling paradigm based on unsupervised learning to enhance spatial diversity in both uplink and downlink. Given users' channel state information (CSI), we first cluster CSIs over the Grassmannian manifold to identify users with high CSI similarity, before scheduling them into MIMO access groups with low co- channel interference. Our paradigm is generalizable to a variety of different simple and scalable unsupervised learning tools and different diversity optimization criteria. Numerical tests demonstrate substantial gain in terms of spectrum efficiency and interference suppression at modest computation complexity.