Daniel L. C. Mack

Integrating Machine Learning Techniques to Adapt Protocols for QoS-enabled Distributed Real-time and Embedded Publish/Subscribe Middleware

Quality-of-service (QoS)-enabled publish/subscribe (pub/sub) middleware provides the infra-structure needed to disseminate data predictably, reliably, and scalably in distributed real-time and embedded (DRE) systems. Maintaining QoS properties as the operating environ¬ment fluctuates is challenging, however, since the chosen mechanism (e.g., transport protocol or caching algorithm for data persistence) may no longer provide the needed QoS. Moreover, some adaptation approaches are tailored for particular types of operating environments, such as environments whose configuration properties (e.g., number of data receivers or data sending rate) are known prior to runtime versus unknown until runtime. For DRE pub/sub systems operating in dynamic environments, adjustments to mechanisms must be timely, accurate for known environments, and resilient to environments unknown until runtime. Several adaptation approaches, such as policy-based [1] and reinforcement learning [2] have been developed to ensure end-to-end quality-of-service (QoS) for enterprise distributed systems in dynamic operating environments. Not all approaches are applicable for DRE pub/sub systems, however, due to their stringent accuracy, timeliness, and development complexity requirements. Supervised machine learning techniques, such as artificial neural networks (ANNs) [3] and support vector machines (SVMs) [4], are promising approaches to address the accuracy, time complexity, and development complexity concerns of adaptive enterprise DRE systems. This article describes the results of research that (1) empirically evaluates supervised machine learning techniques used to adapt the transport protocols of QoS-enabled pub/sub middleware autonomically in a dynamic environment and (2) integrates multiple techniques to increase accuracy for environments known a priori and not known until runtime. Our results show that both ANNs and SVMs provide constant time complexity, low latency, and reduced de-velopment complexity. ANNs are generally more accurate in providing adaptation guidance for environments whose properties are known prior to runtime and provide sub-sec response times, whereas SVMs provide higher accuracy with sec latencies for environments whose properties are not known until runtime. Both approaches can be leveraged together with QoS-enabled pub/sub middleware to address the timeliness, accuracy, and development com-plexity needs of enterprise DRE systems executing in dynamic environments.

Learning Bayesian Network Structures to Augment Aircraft Diagnostic Reference Models

Fault detection and isolation schemes are designed to detect the onset of adverse events during operations of complex systems, such as aircraft and industrial processes. The state-of-the-art fault diagnosis systems on aircraft combine an expert-created reference model of the associations between faults and symptoms, and a Naïve Bayes reasoner. For complex systems with many dependencies between components, the expert-generated reference models are often incomplete, which hinders timely and accurate fault diagnosis. Mining aircraft flight data is a promising approach to finding these missing relations between symptoms and data. However, mining algorithms generate a multitude of relations, and only a small subset of these relations may be useful for improving diagnoser performance. In this paper, we adopt a knowledge engineering approach that combines data mining methods with human expert input to update an existing reference model and improve the overall diagnostic performance. We discuss three case studies to demonstrate the effectiveness of this method.

Using machine learning to maintain pub/sub system QoS in dynamic environments

Quality-of-service (QoS)-enabled publish/subscribe (pub/sub) middleware provides powerful support for scalable data dissemination. It is hard, however, to maintain specified QoS properties (such as reliability and latency) in dynamic environments (such as disaster relief operations or power grids). For example, managing QoS manually is often not feasible in dynamic systems due to (1) slow human response times, (2) the complexity of managing multiple interrelated QoS settings, and (3) the scale of the systems being managed. For certain applications the systems must be able to reflect on the conditions of their environment and adapt accordingly. Machine learning techniques provide a promising adaptation approach to maintaining QoS properties of QoS-enabled pub/sub middleware in dynamic environments. These techniques include decision trees, neural networks, and linear logistic regression classifiers that can be trained on existing data to interpolate and extrapolate for new data. By training the machine learning techniques with system performance metrics in a wide variety of configurations, changes to middleware mechanisms (e.g., associations of publishers and subscribers to transport protocols) can be driven by machine learning to maintain specified QoS. This paper describes how we are applying machine learning techniques to simplify the configuration of QoS-enabled middleware and adaptive transport protocols to maintain specified QoS as systems change dynamically. The results of our work thus far show that decision trees and neural networks can effectively classify the best protocols to use. In particular, decision trees answer questions about which measurements and variables are most important when considering the reliability and latency of pub/sub systems.

A Differential Approach for Identifying Important Student Learning Behavior Patterns with Evolving Usage over Time

Effective design and improvement of dynamic feedback in computer-based learning environments requires the ability to assess the effectiveness of a variety of feedback options, not only in terms of overall performance and learning, but also in terms of more subtle effects on students’ learning behavior and understanding. In this paper, we present a novel interestingness measure, and corresponding data mining and visualization approach, which aids the investigation and understanding of students’ learning behaviors. The presented approach identifies sequential patterns of activity that distinguish groups of students (e.g., groups that received different feedback during extended, complex learning activities) by differences in both total behavior pattern usage and evolution of pattern usage over time. We demonstrate the utility of this technique through application to student learning activity data from a recent experiment with the Betty’s Brain learning environment and four different feedback and learning scaffolding conditions.

Fault Isolation for Spacecraft Systems: An Application to a Power Distribution Testbed

Abstract Modern electrical power disribution systems play a critical role in system operations. Therefore, early fault detection and isolation is essential to maintaining system safety and avoiding catastrophic failures. This paper discusses a fault isolation scheme based on a qualitative fault signature-based isolation mechanism that applies to abrupt, incipient and intermittent faults in the system. We discuss the isolation algorithms for a combination of these faults, and demonstrate their performance on a set of test cases generated from a NASA Ames spacecraft power distribution testbed. Our results show good isolation accuracy with 103 out of 134 faulty scenarios isolated correctly. Most of the isolation errors can be attributed to errors in the detection scheme.