Federated Learning for Multipoint Channel Charting

Abstract

Multipoint channel charting (MP-CC) has been proposed as an effective approach to reap the benefits of cooperation for learning accurate channel charts in massive MIMO systems with multiple bases-stations (BSs). The high-dimensional nature of channel state information (CSI) data, however, imposes significant communication overhead between BSs for joint learning of the MP-CC. To reduce communication overhead and foster locality of CSI data, we explore federated learning (FL) approaches for distributed learning of joint multipoint channel charts. In the proposed approach, each BS learns an individual model and a shared model, where the individual model parameters are unique to each BS and the shared model parameters are communicated to a central server for aggregation. By sharing only weights of the shared model after each training episode, the communication overhead between BSs can be significantly reduced. We provide experimental results on a convolution autoencoder architecture with simulated beam-space CSI data, and compare the FL approach against a fully centralized architecture.