Learning-based Channel State Information Inference

CSI Acquisition: Beyond Pilot-Aided Approach

Channel state information (CSI) plays a vital role in wireless communication systems for both signal processing functions and network operatiaons. However, traditional pilot-aided channel training methods encounter a bottleneck due to network densification and Base Station (BS) sleeping control. Hence, novel channel estimation methods with low pilot overhead are needed.

CSI Acquisition: Inference by Learning

In general, although the CSI of geographically separated BSs is linearly independent, there exist non-linear dependency among them if the BSs are densely deployed and therefore they can be inferred each other. In particular, the CSI of the BSs without pilots can be inferred by the CSI of those BSs with pilots. However, characterizing the non-linear dependency is non-trivial and also needs specific channel models. In the follows, we have incorporated several machine learning approaches to solve the problem.

1. Channel Learning from Centralized CBS

Inspired by the non-linear dependencies among remote channels, we propose a remote beamforming inference scheme which can significantly reduce the pilot resources and training overhead for traffic BSs (TBSs) in hyper-cellular network. The main idea is to infer the optimal beam pattern of TBSs without pilots based on the CSI of centralized control BS (CBS) with pilots. A two hidden-layered neural network is built and trained to make the inference. The simulation results with ray-tracing channel data show that the inference scheme can achieve competitive level of performance as precise-location based approach.

2. Channel Learning from Distributed TBSs^[3][4]

The correlation among channels of adjacent TBSs increases with the density of cells and can be used to infer the channels of sleeping TBSs. Compared to the conventional model-based inference (e.g., Gaussian Process), Neural Network (NN) or K-nearest neighbors (KNN) based channel learning schemes perform much better in practical scenarios. In addition, the prediction mean square error (MSE) can be further reduced by jointly exploiting the information of user location and the channels of adjacent BSs.

[1] J. Liu, R. Deng., S. Zhou, and Z. Niu, “Seeing the unobservable: channel learning for wireless communication networks,” IEEE Global Communications Conference (GLOBECOM), 2015.
[2] R. Deng, Z. Jiang, J. Liu, S. Zhou, and Z. Niu, “CSI acquisition for sleeping cells in hyper cellular networks based on channel learning,” (in Chinese) Sci Sin Inform, 2017, 47: 1583–1591.
[3] S. Chen, Z. Jiang, J. Liu, R. Vannithamby, S. Zhou, Z. Niu, and Y. Wu, “Remote Channel Inference for Beamforming in Ultra-Dense Hyper-Cellular Network,” IEEE Global Communications Conference (GLOBECOM), 2017.