By Periklis Mantenoglou,
National and KApodistrian University of Athen
September 2024
Abstract
Contemporary applications commonly demand the detection of ‘situations of interest’ in real-time and with minimal latency. In maritime situational awareness, e.g., it is crucial to identify and report vessel behaviours that may indicate dangerous, illegal or environmentally hazardous activities in a timely manner. Monitoring situations of interest is challenging, as large, high-velocity streams of data are being generated continuously, and it is not feasible to store these streams in memory and process them at a later time. Stream reasoning involves the real-time detection of critical situations by reasoning over large volumes of incrementally-arriving, symbolic data, typically termed as ‘events’. In order to support contemporary applications, a stream reasoning system needs to employ efficient reasoning algorithms that are tailored for the streaming setting, a highly expressive pattern specification language with a formal and declarative semantics, as well as techniques for handling noise and uncertainty.
This thesis is motivated by the absence of a framework with all the aforementioned features, and aims at extending state-of-the-art computational frameworks in order to support stream reasoning effectively. We focus on frameworks that employ the Event Calculus, a logic programming formalism for representing events and reasoning about their effects over time. The Event Calculus exhibits a formal, declarative semantics, while supporting both instantaneous and durative events, hierarchical and relational event specifications, temporal persistence and background knowledge. RTEC is a computational framework that is based on the Event Calculus and includes several optimisation techniques for stream reasoning.
We propose three extensions of RTEC, i.e., RTEC◦, RTEC→ and RTECA, supporting, respectively, temporal specifications with cyclic dependencies, events with delayed effects and situations specified using the relations of Allen’s interval algebra. Moreover, we study PIEC, i.e., an Event Calculus-based system for probabilistic event recognition, and present oPIEC, i.e., an extension of PIEC for reasoning over noisy event streams. In order to support large data streams, we propose two bounded-memory versions of oPIEC. All of our proposed frameworks have a formal semantics and feature optimised stream reasoning algorithms that have been assessed both theoretically and empirically, using large, real and synthetic data streams, in experiments that included comparisons with state-of-the-art frameworks. Our analysis demonstrates that RTEC◦, RTEC→ and RTECA are suitable for stream reasoning with complex temporal specifications, while the variants of oPIEC exhibit high predictive accuracy and reasoning efficiency when reasoning over noisy data streams.