Showing papers by "Alexander Artikis published in 2022"

Stream Reasoning with Cycles

Abstract: Temporal specifications, such as those found in multi-agent systems, often include cyclic dependencies. Moreover, there is an increasing need to evaluate such specifications in an online manner, upon streaming data. Consider, for example, the online computation of the normative positions of the agents engaging in an e-commerce protocol. We present a formal computational framework that deals with cyclic dependencies in an efficient way. Moreover, we demonstrate the effectiveness of our framework on large synthetic and real data streams, from the fields of multi-agent systems and composite event recognition.

Online fleet monitoring with scalable event recognition and forecasting

Abstract: Moving object monitoring is becoming essential for companies and organizations that need to manage thousands or even millions of commercial vehicles or vessels, detect dangerous situations (e.g., collisions or malfunctions) and optimize their behavior. It is a task that must be executed in real-time, reporting any such situations or opportunities as soon as they appear. Given the growing sizes of fleets worldwide, a monitoring system must be highly efficient and scalable. It is becoming an increasingly common requirement that such monitoring systems should be able to automatically detect complex situations, possibly involving multiple moving objects and requiring extensive background knowledge. Building a monitoring system that is both expressive and scalable is a significant challenge. Typically, the more expressive a system is, the less flexible it becomes in terms of its parallelization potential. We present a system that strikes a balance between expressiveness and scalability. Going beyond event detection, we also present an approach towards event forecasting. We show how event patterns may be given a probabilistic description so that our system can forecast when a complex event is expected to occur. Our proposed system employs a formalism that allows analysts to define complex patterns in a user-friendly manner while maintaining unambiguous semantics and avoiding ad hoc constructs. At the same time, depending on the problem at hand, it can employ different parallelization strategies in order to address the issue of scalability. It can also employ different training strategies in order to fine-tune the probabilistic models constructed for event forecasting. Our experimental results show that our system can detect complex patterns over moving entities with minimal latency, even when the load on our system surpasses what is to be realistically expected in real-world scenarios.

Early Time-Series Classification Algorithms: An Empirical Comparison

Abstract: Early Time-Series Classification (ETSC) is the task of predicting the class of incoming time-series by observing as few measurements as possible. Such methods can be employed to obtain classification forecasts in many time-critical applications. However, available techniques are not equally suitable for every problem, since differentiations in the data characteristics can impact algorithm performance in terms of earliness, accuracy, F1-score, and training time. We evaluate six existing ETSC algorithms on publicly available data, as well as on two newly introduced datasets originating from the life sciences and maritime domains. Our goal is to provide a framework for the evaluation and comparison of ETSC algorithms and to obtain intuition on how such approaches perform on real-life applications. The presented framework may also serve as a benchmark for new related techniques.

Resource-effective exploration of tumor treatments with multi-scale simulations

Abstract: Machine learning is regularly used to interpret and analyze information from large and complex datasets originating from numerous fields. In Bioinformatics, the exploration of potentially beneficial drug configurations for tumor treatments via simulations requires multiple processing units to be used in parallel and a considerable amount of time to be completed. In this paper, we apply a state-of-the-art model exploration workflow for the characterization of a new drug configuration parameter space, using a redesigned simulator. Moreover, we incorporate different clustering and optimization approaches and compare their performance in in-silico simulation trials on high-performance computing infrastructure, with respect to time and resource efficiency. The overall goal is to discover regions in this parameter space that can lead to more viable treatments in reasonable time, and thus guide the related research towards more focused and effective real-world trials.

Optimizing complex event forecasting

Abstract: In Complex Event Recognition (CER), applications express business rules in the form of patterns and deploy them in a CER Engine which seeks the occurrence of such patterns on incoming streams. This is useful for practical applications which rely on the timely detection of patterns to support critical decisions. One step further, stakeholders want to act proactively, accurately forecasting the occurrence of patterns on raw streams well ahead of time to better schedule their decisions. This calls for making the transition from CER to Complex Event Forecasting (CEF). In CEF, stochastic models of future behavior are embedded into the event processing loop to project into the future the sequence of events that have occurred so far and to estimate the likelihood of the imminent occurrence of more complex patterns. CEF performance engages the stochastic model's training speed and forecast accuracy. In turn, these performance dimensions are affected by few parameters. However, CEF parameter tuning so that optimal CEF performance is achieved is a non-trivial task. This is due to the fact that there is an infinite number of possible parameter combinations, each affecting CEF performance in ways which are hard to predict. In this work, we introduce the first CEF Optimizer that gracefully automates CEF parameter tuning decisions, rapidly cherry picking good CEF configurations. We detail the novel internal architecture of our CEF Optimizer and present an elaborate empirical analysis on two applications that illustrates the effectiveness of our optimization approach.