Design of Early Reliability Prediction Models for Software Projects

Abstract

The significance of software in modern civilization spans across social, political, fi nancial, healthcare, and military domains. However, the increasing complexity and size of software pose challenges, including jeopardizing quality and driving up testing costs. Software reliability is crucial, especially for mission-critical and high-assurance applica tions. This thesis aims to enhance software reliability by predicting faulty modules and estimating development efforts early in the software development lifecycle. Existing pre diction models fall into two categories: Software Reliability Growth Models (SRGMs) and Early Software Reliability Prediction (ESRP) Models. While reliability growth models of fer estimates, relying solely on them for corrective actions can be costly and delayed. Early prediction facilitates refined project planning, timely delivery, and cost overruns, mitigates overestimation and underestimation, allocate resources effectively, and formulates the op timal development approaches. Software fault prediction (SFP), including Within-Project Fault Prediction (WPFP) and Cross-Project Fault Prediction (CPFP) models, improves re liability. Additionally, effort estimation reduces cost estimation uncertainty and enhances software quality by estimating required manpower early in development. The main objectives of this thesis include: (i) To predict the software fault-prone mod ules in within-project through the Weighted Average Centroid based Imbalance Learning approach. (ii) To predict the software fault-prone modules in a cross-project through simi larity based source project and training data selection techniques. (iii) To predict the soft ware fault-prone modules in a cross-project using applicability based source project selec tion, resampling, and feature reduction. (iv) To estimate the software development effort for a project through a two-stage optimization technique. Firstly, a diverse imbalance learning technique is designed for WPFP. The prediction performance is enhanced by diverse synthetic data generation and noisy data elimination. Secondly, this study utilizes the Wilcoxon Signed Rank (WSR) test to identify similar source projects for cross-project prediction. A novel oversampling technique is introduced to address distribution gaps and skewed distributions. Additionally, the performance of CPFP is improved through the use of the Binary-RAO algorithm. This algorithm explores iii diverse combinations of software features and hyperparameters within a specified search space to extract highly correlated features related to module faultiness. Thirdly, the second objective is extended by integrating applicability scores with similarity scores to improve the accuracy of source project selection. A novel resampling technique is devised to extract highly correlated and similar instances while discarding irrelevant ones from the source data, thus enhancing the efficiency of training data construction and reducing distribution discrepancies and class imbalances. Additionally, to tackle the high-dimensionality issue, an efficient deep learning-based Stacked Autoencoder (SAE) model is developed for fea ture reduction, leading to enhanced performance in CPFP. Lastly, an Adaptive Neuro-Fuzzy Inference System (ANFIS) estimation model is developed using multi-objective optimiza tion techniques for efficient software development effort estimation. The multi-objective rank-based improved Social Network Search (SNS) algorithm is applied to extract optimal software project features and to identify ANFIS parameters. The experimental results of this research demonstrate the importance of imbalance learning, source selection, instance selection, and feature optimization for software datasets and the effectiveness of metaheuristic optimization techniques in effort estimation models. The proposed methods outperform existing methods, offering improved software fault pre diction and effort estimation performance, thereby improving reliability prediction overall.