Learning-based communication system design – autoencoder for (differential) block coded modulation designs and path loss predictions

Abstract

Shannon’s channel coding theorem states the existence of long random codes that can make the error probability arbitrarily small. Recently, advanced error-correcting codes such as turbo and low-density parity-check codes have almost reached the theoretical Shannon limit for binary additive white Gaussian noise channels. However, designing optimal high-rate short-block codes with automatic bit-labeling for various wireless networks is still an unsolved problem. Deep-learning-based autoencoders (AE) have appeared as a potential near-optimal solution for designing wireless communications systems. We take a holistic approach that jointly optimizes all the components of the communication networks by performing data-driven end-to-end learning of the neural network-based transmitter and receiver together. Specifically, to tackle the fading channels, we show that AE frameworks can perform near-optimal block coded-modulation (BCM) and differential BCM (d-BCM) designs in the presence and absence of the channel state information knowledge. Moreover, we focus on AE-based designing of high-rate short block codes with automatic bit-labeling that are capable of outperforming conventional networks with larger margins as the rate R increases. We also investigate the BCM and d-BCM from an information-theoretic perspective. With the advent of internet-of-things (IoT) networks and the widespread use of small devices, we face the challenge of limited available bandwidth. Therefore, novel techniques need to be utilized, such as full-duplex (FD) mode transmission reception at the base station for the full utilization of the spectrum, and non-orthogonal multiple access (NOMA) at the user-end for serving multiple IoT devices while fulfilling their quality-of-service requirement. Furthermore, the deployment of relay nodes will play a pivotal role in improving network coverage, reliability, and spectral efficiency for the future 5G networks. Thus, we design and develop novel end-to-end-learning-based AE frameworks for BCM and d-BCM in various scenarios such as amplify-and-forward and decode-and-forward relaying networks, FD relaying networks, and multi-user downlink networks. We focus on interpretability and understand the AE-based BCM and d-BCM from an information-theoretic perspective, such as the AE’s estimated mutual information, convergence, loss optimization, and training principles. We also determine the distinct properties of AE-based (differential) coded-modulation designs in higher-dimensional space. Moreover, we also studied the reproducibility of the trained AE framework. In contrast, large bandwidth and worldwide spectrum availability at mm-wave bands have also shown a great potential for 5G and beyond, but the high path loss (PL) and significant scattering/absorption loss make the signal propagation challenging. Highly accurate PL prediction is fundamental for mm-wave network planning and optimization, whereas existing methods such as slope-intercept models and ray tracing fall short in capturing the large street-by-street variation seen in urban cities. We also exploited the potential benefits of AE framework-based compression capabilities in mm-wave PL prediction. Specifically, we employ extensive 28 GHz measurements from Manhattan Street canyons and model the street clutters via a LiDAR point cloud dataset and 3D-buildings by a mesh-grid building dataset. We aggressively compress 3D-building shape information using convolutional-AE frameworks to reduce overfitting and propose a machine learning (ML)-based PL prediction model for mm-wave propagation.