Semi-Supervised Learning for MIMO Detection
Peiyan Ao, Runhua Li, Rongchao Sun, Jiang Xue · 2022
The model-driven deep learning method has been verified to be effective for signal detection in the massive multi-input multi-output (MIMO) system. In previous work, this kind of methods need to be trained with numerous pilots in a supervised manner, which will occupy amount of spectrum resources. In addition, the abundant information in the symbols are not utilized in the training procedure. To solve this issue, Firstly in this paper, a new unsupervised deep learning (DL) network named Un-OAMPNet is proposed, which considers Mixture of Gaussian (MoG) noise model under a maximum a posterior (MAP) framework. Secondly, Un-OAMPNet is extended to the semi-supervised DL network (Semi-OAMPNet) with a few pilots to increase the detection performance. In Semi-OAMPNet, the loss function is combined with the mean square error (MSE) loss in the supervised manner and the MAP loss in the unsupervised manner. In this way, Semi-OAMPNet inherits the advantages of OAMP-Net2 and gains better detection performance with fewer pilots. Simulation results show that the Un-OAMPNet is effective without any pilots in the training procedure and proposed Semi-OAMPNet has better performance compared with other model-driven detectors.