Bit-Level Optimized Neural Network for Multi-Antenna Channel Quantization

Abstract

Quantized channel state information (CSI) plays a critical role in precoding design which helps reap the merits of multiple-input multiple-output (MIMO) technology. In order to reduce the overhead of CSI feedback, we propose a deep learning based CSI quantization method by developing a joint convolutional residual network (JC-ResNet) which benefits MIMO channel feature extraction and recovery from the perspective of bit-level quantization performance. Experiments show that our proposed method substantially improves the performance. Index TermsChannel state information (CSI), quantization, neural network (NN), multiple-input multiple-output (MIMO). I. INTRODUCTIONMultiple-input multiple-output (MIMO) technology has shown its ability in obtaining rich diversity and multiplexing gains. With recent development of MIMO, the trend of deploying more antennas comes with a number of challenges. One of the major challenges is to acquire accurate high-dimensional MIMO channel state information (CSI) at the transmitter, especially in frequency division duplex (FDD) systems.An increasing number of antennas results in a high-dimensional channel matrix, which makes it difficult for the transmitter to obtain accurate CSI through a feedback channel of limited bandwidth.Conventional codebook based methods quantize the channel matrix into a sequence of bits which represents the index of a codeword [1]. Considering that the complexity of the codebook based methods grows exponentially with the number of quantization bits, it can be prohibitively difficult for large MIMO channels. Compressed sensing (CS) based methods, e.g., TVAL3 [2], can be used for solving this difficulty Manuscript ). 2 through feature extraction and dimension compression. However, when the compression ratio is extremely high, the performance of these CS based methods degrades severely [3].Recently, neural network (NN) based methods, inspired from CS based methods, showed appealing performance for MIMO CSI compression [3][4][5]. Previously, encoder-decoder schemes were proposed in [6,7] to learn the behaviors of a transmitter in wireless communication systems. This idea of an encoderdecoder network was then utilized to learn the MIMO CSI compression [3][4][5]. In particular, the encoder network compressed the channel matrix into a low dimensional vector while the decoder network was responsible for recovering the channel matrix directly from the compressed vector. Usually the decoder was designed with a more complex structure than the encoder because the base station (BS) generally has stronger computing ability than user equipment (UE).The NN based method in [3], namely CsiNet, assumed that the decoder achieves a compressed CSI vector of continuous values. Then in [4,5], the CsiNet was enhanced by utilizing a long-short time memory (LSTM) network [8] to further exploit temporal channel correlations. In a digital communication system with limited bandwidth, however, it is still impossible to transmit a vector of continuous values even though it is highly compresse...

ELM-Based Superimposed CSI Feedback for FDD Massive MIMO System

“…In general, only the real-valued data can be processed by DL-based network, e.g., [7]- [15], [24], etc. Unlike the DL-based network, ELM network can directly process complex-valued data [28], avoiding the conversion from complex-valued data to real-valued data.…”

Power Consumption Optimization Using Gradient Boosting Aided Deep Q-Network in C-RANs

“…Recent works have achieved significant improvement in computational efficiency by applying the DNN approximator [17] [18] [19] [20] to DRA problems. However, the unstable performance of DNN in regression process makes it hard to achieve good performance [21].…”

A Novel Cross Entropy Approach for Offloading Learning in Mobile Edge Computing

“…Furthermore, a deep Q-network was utilized in [5] to optimize the computational offloading, without a priori knowledge of the network. Most of these methods, including those which use deep learning network (DNN), focused on the offloading design from a perspective of long-term optimization and at the cost of complexity and robustness [6] [7]. Moreover, these methods can hardly track fast channel changes, due to the requirement of offline learning.…”