Το έργο με τίτλο Ομαδοποίηση αλγορίθμων συμπερασμού σε δίκτυα επικοινωνιών από τον/τους δημιουργό/ούς Kariotakis Emmanouil διατίθεται με την άδεια Creative Commons Αναφορά Δημιουργού 4.0 Διεθνές
Βιβλιογραφική Αναφορά
Εμμανουήλ Καριωτάκης, "Ομαδοποίηση αλγορίθμων συμπερασμού σε δίκτυα επικοινωνιών", Διπλωματική Εργασία, Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών, Πολυτεχνείο Κρήτης, Χανιά, Ελλάς, 2022
https://doi.org/10.26233/heallink.tuc.93442
This work offers an algorithmic framework for in-network inference, using message passing among ambiently powered wireless sensor network (WSN) terminals. The stochastic nature of ambient energy harvesting dictates intermittent operation of each WSN terminal and as such, the message passing inference algorithms should be robust to asynchronous operation. A version of Gaussian Belief Algorithm (GBP) is described, which can be reduced to an affine fixed point (AFP) problem, used to solve linear systems of equations. To achieve this, we have to cluster the Probabilistic Graphical Model (PGM) behind GBP, in order to map it to the WSN terminals. We propose two different clustering approaches, namely edge and node clustering. For the first approach, we explain the reasons why a previous method does not produce the expected results and we offer another method, which performs better. We also explain limitations of edge-based clustering. On the other hand, node clustering has a clear metric for performance, which is relevant to the number of edges connecting the different clusters. For this approach, we utilize three different clustering algorithms, the k-means, the spectral clustering and an autonomous, in-network clustering algorithm. Furthermore, we show in both theory and simulation that there is strong connection between spectral radius and the convergence rate of AFP problems with probabilistic asynchronous scheduling. The latter corroborates known theory for synchronous scheduling. Interestingly, it is shown through simulations that different clustering offers similar convergence rate, when probabilistic asynchronous scheduling is utilized with carefully selected probabilities that accelerate convergence rate in the mean sense. Finally, we show an existing distinction between convergence rate and energy consumption of the network and we present experimental results comparing the different clustering methods. In most cases, spectral clustering outperforms the rest, with reduced energy consumption (by a factor of 2 compared to k-means in specific cases).