Unsupervised Learning for Resource Allocation and User Scheduling in Wideband MU-MIMO Systems

• Published in: IEEE open journal of the Communications Society Vol. 5; p. 1
• Main Authors: Hsu, Chih-Ho, Ding, Zhi
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Broadband; Clustering; Clustering algorithms; Co-Channel Interference (CCI); Cochannel interference; Correlation; graph embedding; Interchannel interference; MIMO communication; MultiUser MIMO (MU-MIMO); Resource allocation; Resource scheduling; Scheduling; Signal to noise ratio; Unsupervised learning; user grouping; User groups; Wideband; wideband user scheduling
• ISSN: 2644-125X; 2644-125X
• DOI: 10.1109/OJCOMS.2024.3384110

By leveraging spatial diversity, MultiUser MIMO (MU-MIMO) can serve multiple users over shared time-frequency Resource Blocks (RBs) and substantially improve spectral efficiency. However, performances of wideband MU-MIMO systems are severely limited by both frequency-selective channels and Co-Channel Interference (CCI) among users. To reach the full potential of MU-MIMO, users should be scheduled at RBs with decent channel gains and minimal CCIs. Since such scheduling problem is NP-hard and the transmission time interval of modern wireless systems is ultra-short, it is critical to design efficient algorithms that can make satisfactory sub-optimal user scheduling decisions in real-time. Nonetheless, existing works either rely on heuristics or may not readily be applied to wideband system. To tackle these challenges, we propose a novel Unsupervised Learning-Aided Wideband Scheduling (ULAWS) framework. Specifically, ULAWS first utilizes Multi-Dimensional Scaling (MDS) based graph embedding and clustering to obtain intrinsic user groups with low CCI among co-channel users. Based on clustering results, we adopt Gale-Sharpley algorithm to find a stable matching between users and RBs. Next, a graph-based post-processing procedure stacked with three efficient steps is applied as refinement. Simulation results demonstrate performance gain over benchmark methods in terms of sum rate, fairness and outage rate, under various system parameters and scenarios.