Throughput Optimization in a Wireless Communication Channel Using Machine Learning Algorithms

Conventional models employed for the physical layer of wireless telecommunication systems often rely on the theoretically proven optimal mathematical models. In practice, their implementations can be affected by various non-linearities or hardware limitations (e.g. non-linearity in power amplifiers). As a result, solutions based on conventional architectures can be suboptimal due to the fact that mathematically tractable models can capture technical imperfections only approximately. In other words, their rigid analytical structure may lead to reduced responsiveness as wireless system performance changes or a real channel environment evolves in time. Due to this, performance can deteriorate for implementations involving higher bit-rate transmission such as 64-QAM or 256-QAM. It is here that end-to end trainable communication systems can be utilized to adress these challenges for an improved end-to-end performance. Within this context, autoencoders based on a deep neural network (DNN) can be used to implement trainable constellations and neural demappers (ND) that can be trained in an end-to-end manner to obtain optimized constellation geometries responsive to wireless environments with higher data rates. This thesis explores the performance of four autoencoder-based architectures that incorporate a trainable constellation and a neural demapper. These architectures are compared against a baseline model that consists of a conventional, non-trainable 64-QAM constellation and a conventional max-log demapper. For the purpose of providing a more general frame of reference, the proposed simulations also include a conventional analytical model consisting of a non-trainable 64-QAM constellation and an analytical demapper. Based on the simulations, the autoencoder-based architectures outperformed both the baseline and analytical models for the 64-QAM setting, achieving a better singal-to-noise ratio (SNR) performance of 0.5 dB and delivering a relative decrease in block error rate (BLER) of up to 80% with respect to to the conventional models. Keywords: autoencoder, end-to-end trainable system, wireless communications, physical link, trainable constellation, neural demapper, analytical model