Optimization based machine learning algorithms for software reliability growth models

Abstract

Software reliability is a critical factor for system performance and safety, especially in defense industries, where operational failures can have severe consequences. To evaluate and improve software reliability, Software Reliability Growth Models (SRGMs) are widely used. However, many previous studies have relied on single optimization methods or deep learning approaches, which are prone to local optima and extrapolation issues, reducing prediction accuracy. To fill this gap, current study employs a broader range of optimization algorithms based on the Least Squares Method (LSM) and Maximum Likelihood Estimation (MLE) to approximate global optima. NASA’s Jet Propulsion Laboratory (JPL) software defect datasets were used, and several widely recognized SRGM models, including Goel-Okumoto, Delayed S-Shape, Inflection S-Shape, Weibull, and Log-Logistic, were evaluated. Experimental results show that the choice of optimization method significantly affects prediction performance, as measured by Mean Squared Error (MSE). For example, in the J2 dataset, the Weibull model exhibited MSE values ranging from 70.778 to 15,767.68—a 222-fold difference—demonstrating the critical role of optimization in prediction accuracy. The findings confirm the risks of relying solely on single-method approaches and highlight the value of diverse optimization strategies for achieving near-global optima. The study presents a practical framework for improving software reliability assessments, contributing to the development of highly reliable software for the defense industry.