Unlabeled sensing and detection for improved energy efficiency in wireless sensor networks

Abstract

While optimal solutions to the classical problem of linear regression have been thoroughly studied, the relatively new problem of shuffled linear regression, or unlabeled sensing, presents an open challenge. The goal of unlabeled sensing is to estimate measurements which have been linearly transformed and permuted. A closely related problem is that of unlabeled detection, which aims to perform global detection of some phenomenon based on local measurements or decisions that have been permuted. In both cases, the permutation of the observations is unknown to the observer. These unlabeled methods have important applications, such as energy-efficient communication techniques within a wireless sensor network, which is the motivating idea behind this work. Hence, this thesis aims to develop new solutions to the problems of unlabeled sensing and detection and provide insights into their effectiveness. First, a novel deep learning approach is explored to provide an efficient and accurate estimation technique for the general unlabeled sensing problem, which is shown to greatly improve computational efficiency with very little deterioration of recovery performance. Next, a heuristic algorithm which combines the ideas of compressive sensing and deep learning is developed to address the unlabeled sensing problem in a heterogeneous wireless sensor network. The performance of this algorithm is then tested and compared to that of the current state-of-the-art approach, demonstrating improved performance. Lastly, the problem of unlabeled detection in a similar network is studied. An analytical detector based on the generalized likelihood ratio test and maximum likelihood estimation is derived, as well as a heuristic detector based on deep learning techniques. These detectors are then simulated and compared, with both demonstrating impressive performance under various conditions.