Machine Learning for Reliable Communication and Improved Tracking of Dynamical Systems

Vedula, Kirty P.

Etd

Machine Learning for Reliable Communication and Improved Tracking of Dynamical Systems

Public

Reliable communication systems and optimal tracking of dynamic systems are subjects that have been studied for several decades. In recent years, however, there is a renewed interest in these subjects from the perspective of machine learning. This dissertation applies machine learning techniques to develop new insights and specific algorithms for two important problems: (i) achieving reliable communications in noisy channels using autoencoders and (ii) improving tracking of dynamical systems using machine learning. The first problem considers a joint coding and modulation scheme for an end to-end communication system design using an autoencoder architecture in short blocklength regime. Unlike the classical approach of separately designing error correction codes and modulation schemes for a given channel, the approach here is to learn an optimal mapping directly from messages to channel inputs while simultaneously learning an optimal mapping directly from channel outputs to estimated messages. Additive white Gaussian noise (AWGN) channels are considered first and the performance of the autoencoder is compared against various coding and modulation schemes for linear block codes. An analysis on conditional block error rate is presented with interesting insights on the geometry of the codewords generated by the autoencoder. For AWGN channels, numerical results show that the autoencoder can achieve better block error rate (BLER) performance than BPSK modulated Hamming codes with maximum likelihood decoding. We extend it to other non-canonical channels which have no-known good codes such as Bernoulli-Gaussian Impulsive Noise (BGIN) channel. A family of autoencoders is developed for different probabilistic parameters in BGIN channels. Numerical results show the autoencoder achieves uniformly better BLER performance than conventional block codes. The proposed architecture is general and can be modified for comparison against other block coding schemes and higher-order modulations. The second problem considers tracking dynamical systems under parametric mismatch and model switching. The Kalman filter (KF) is the optimal state estimator for linear and Gaussian dynamic systems. Optimality of the KF, however, requires exact knowledge of covariances and the system dynamics. Otherwise, the KF will be ``mismatched'' and, consequently, the state estimates and predictions will not be optimal. We develop a machine learning approach to accurately predict the dynamic states without knowledge of the system dynamics or noise statistics. Considering an application of oscillator phase predictions, we demonstrate that the machine learning approach achieves performance close to optimal irrespective of the amount of parametric mismatch. Then, model switching is considered in the context of (i) maneuvering target-tracking and (ii) tracking dynamical systems over Gilbert-Elliott Channels. An Interacting Multiple Model (IMM) is a commonly used state estimator that utilizes hypotheses from a filter bank of KFs instead of a single KF to handle changeable and uncertain maneuvering movements. We develop an alternate machine learning method using a Temporal Convolutional Network (TCN) to demonstrate better stability robustness to model switching. We also develop a hybrid model, Autoencoder Interacting Multiple Model (AEIMM) filter, as an extension to Autoencoder Kalman Filter (AEKF). AEIMM embeds an IMM within an autoencoder framework to learn measurements and their associated measurement covariances for multiple dynamic models. We demonstrate that AEIMM outperforms state-of-the-art maneuvering target-tracking algorithms such as IMM and machine learning models like Long-Short Term Memory Networks.

Creator