CONFIRMED: EAS Doctoral Proposal Defense by Fatemeh Salboukh
Date: Tuesday, January 21, 2025 Time: 10am Topic: Statistical and Machine Learning Models for System Reliability and Resilience Location: TXT 105 (CSCDR) Abstract: Software reliability and resilience are essential for ensuring dependable system performance, particularly in the face of evolving demands and unexpected disruptions. Traditional reliability models, such as the Non-Homogeneous Poisson Process (NHPP), have been widely used to predict defect occurrence based on testing time or effort. However, these models often fall short of capturing the complexities of real-world systems. Resilience engineering, which focuses on a system's ability to respond to and recover from shocks, has gained significant attention as a complementary approach to reliability. While statistical models provide foundational insights, their rigid assumptions limit flexibility and fail to account for dynamic patterns in defect occurrence and recovery processes. On the other hand, machine learning methods, such as neural networks, offer the potential to model intricate dependencies and non-linear trends. However, these models often require extensive data, which is not always available in resilience engineering contexts, and may lack robustness in long-term predictions. This gap highlights the need for integrated approaches that effectively address the challenges of modeling resilience in systems experiencing varying types and intensities of shocks. In order to address these challenges, this dissertation proposes hybrid approaches including: (i) For defect prediction, we use recurrent neural networks (RNNs) and long short-term memory (LSTM) networks that incorporate covariates to improve predictive accuracy by reflecting key factors driving defect discovery, (ii) To enhance defect classification, we apply Locally Linear Embedding (LLE) as a preprocessing step, which simplifies complex data, allowing classifiers to better interpret defect patterns, and (iii) For resilience assessment, we introduce a transfer function model, a flexible time series approach that considers multiple stressors and recovery patterns. This model captures the dynamic response of a system under varied shocks, allowing for a more accurate resilience evaluation without needing extensive data. By combining machine learning and statistical methods, this dissertation aims to advance both reliability and resilience assessment in software systems, providing robust, adaptable models capable of predicting defects and tracking recovery under complex conditions. These contributions support the development of systems that not only maintain performance but also adapt to future challenges with resilience. ADVISOR(S): Dr. Lance Fiondella, Department of Electrical & Computer Engineering (lfiondella@umassd.edu) COMMITTEE MEMBERS: Dr. Alfa Heryudono, Department of Mathematics Dr. Ruolin Zhou, Department of Electrical & Computer Engineering Dr. Hong Liu, Department of Electrical & Computer Engineering NOTE: All EAS Students are ENCOURAGED to attend.
Textiles Building 105
Engineering and Applied Sciences
easphd@umassd.edu