Massive MIMO CSI Feedback using Channel Prediction: How to Avoid Machine Learning at UE?

TL;DR

A novel hybrid framework composed of a recurrent neural network and NP, which yields better prediction accuracy than individual models, is proposed, which is inspired to use a statistical model and ML together.

Abstract

In the literature, machine learning (ML) has been implemented at the base station (BS) and user equipment (UE) to improve the precision of downlink channel state information (CSI). However, ML implementation at the UE can be infeasible for various reasons, such as UE power consumption. Motivated by this issue, we propose a CSI learning mechanism at BS, called CSILaBS, to avoid ML at UE. To this end, by exploiting channel predictor (CP) at BS, a light-weight predictor function (PF) is considered for feedback evaluation at the UE. CSILaBS reduces over-the-air feedback overhead, improves CSI quality, and lowers the computation cost of UE. Besides, in a multiuser environment, we propose various mechanisms to select the feedback by exploiting PF while aiming to improve CSI accuracy. We also address various ML-based CPs, such as NeuralProphet (NP), an ML-inspired statistical algorithm. Furthermore, inspired to use a statistical model and ML together, we propose a novel hybrid framework composed of a recurrent neural network and NP, which yields better prediction accuracy than individual models. The performance of CSILaBS is evaluated through an empirical dataset recorded at Nokia Bell-Labs. The outcomes show that ML elimination at UE can retain performance gains, for example, precoding quality.

Figures (11)

• Figure 1: Graphical illustration of time-series NNs, i.e., LSTM and BiLSTM. Fig. $1$ (a) shows single-cell of an LSTM, where input is real-valued CSI realization for time instant $t$. Fig. $1$ (b) depicts a fully connected BiLSTM-based NN, where inputs are learned in two ways. Circular shapes given in Fig. $1$ (b) denote a single-cell of LSTM, which is given in Fig. $1$ (a). The outputs of BiLSTM portray predicted CSI realizations. Source: karam_CP_Globecom.
• Figure 2: Flow diagram of hybrid model, where dotted boxes are showing the prediction models used for the hybrid approach. Processed data is the real-valued CSI realizations, and final prediction denotes the multi-step ahead predicted CSI realizations. Source: karam_CP_Globecom.
• Figure 3: Pictorial representation of two feedback resources in probabilistic and error-bin feedback. The former utilizes both resources for CSI feedback and later uses one to check contention and the second to feedback CSI of UE winning the contention.
• Figure 4: View of Nokia Bell-Labs campus in Stuttgart, Germany, where measurements are recorded. A mMIMO antenna, represented with a blue bar on extreme left-side of the figure, is located on a rooftop, transmitting in the direction represented with arrow. The UE is moved on the different tracks, illustrated with black lines, where arrowhead representing the direction of UE's movement and numerical values showing the track number. Source: Karam_CP.
• Figure 5: A closer view of measurement track-1 and network entities is shown in these figures. Buildings of approximately $15$ m height can also be seen, which are acting as reflectors and blockers for the radio waves. Source: CSI_TCCN_Karam.