נושא הפרוייקט

מספר פרוייקט מחלקה שמות סטודנטים אימייל שמות מנחים

עימעום ופיצול באלגוריתמי היסק מבוזרים לא שלמים. למה זה עובד?

Damping and splitting in incomplete distributed inference algorithms. Why does it work?

תקציר בעיברית תקציר באנגלית

Distributed inference algorithms are widely used in various domains, and two techniques that have shown promise in improving their performance are Damping and Splitting. This research project aims to investigate why and how Damping and splitting work together to enhance the performance of incomplete distributed inference algorithms.

Initially, our focus was to comprehend the individual impact of Damping on these algorithms. Subsequently, we will examine the isolated effect of Splitting. Once we obtain a thorough understanding of each technique, we will proceed to explore their combined effects. We anticipate that comprehending these techniques individually will provide valuable insights into their joint application. To approach this study, we divided it into three distinct stages: analyzing a chain, followed by a single cycle, and concluding with a multiple-cycle graph. This sequential progression aligns with established practices when studying such algorithms. 
In our investigation of Damping within a chain, we successfully deduced the convergence behavior of the exchanged messages. Consequently, we derived a formula for calculating the messages exchanged by each agent in a chain employing Damping. Furthermore, we achieved a significant milestone by describing the structure of the Backtrack Cost Tree (BCT) associated with Damping in a single cycle. This achievement is noteworthy as previous literature assumed that the BCT of Damping remains unchanged. 
To summarize, we have gained a comprehensive understanding of the effects of Damping in a chain and within a single cycle. Our subsequent step involves exploring its behavior in a multi-cycle environment, with the aim of concluding the individual analysis of Damping and transitioning to the examination of Splitting.